newsletterii.tudelft.nl/bnvki/wp-content/uploads/2017/03/27.5.pdf · there is a famous conjecture,...

NE

WS

LE

TT

ER

News from theBenelux Associationfor Artificial Intelligence

October 2010Vol. 27, No. 5ISSN 1566-8266

Ramanujaniana (part 1)

The Magic of Monte-Carlo Tree Search

The 2010 SKBS-Strukton Prize

BNVKI Newsletter October 2010106

1729

Editor-in-chief

Srinivasa Iyengar Ramanujan (December 22, 1887 – April 26, 1920), or Ramanujan for short, was an Indianmathematical genius who, with almost no formal training in pure mathematics, made substantial contributions tomathematical analysis, number theory, infinite series and continued fractions. Ramanujan’s talent was said, bythe prominent English mathematician G.H. Hardy, to be in the same league as legendary mathematicians such asEuler, Gauss, Newton and Archimedes. Ramanujan was fond of infinite series. In particular, he discovered anexpression for , enabling the quick calculation of successive decimals of . By this formula, has at the end ofthe 1980s been calculated up to over a billion digits.

( )∑∞

=

+=

044 )396(

]263901103[)!(!4

980181

nn

nn

nπ

Although he died at the very young age of only 32, his achievements are overwhelming. As a consequence, hisname is appearing in many fields in mathematics. For example, there are prime numbers obeying some specifiedconstraints, now known as Ramajunan primes. There exists a generalization of the Jacobi theta function, nowcalled the Ramanujan theta function. There is a famous conjecture, of course called the Ramanujan conjecture,relating Fourier coefficients and prime numbers. His name even can be found in relation with the golden ratio

=2

51+ ), for which Ramanujan found the following beautiful expression:

Also, there is a Ramanujan Prize, honouring a researcher, younger than 45 years old, who has conductedoutstanding mathematics research in a developing country. There even is a mathematical journal, entitled TheRamanujan Journal, which is completely devoted to areas of mathematics influenced by his work.

A common anecdote about Ramanujan relates to the number 1729. Hardy arrived at Ramanujan’s residence in acab numbered 1729. Hardy commented that the number 1729 seemed to be uninteresting. Ramanujan is said tohave stated on the spot that it was actually a very interesting number mathematically, being the smallest naturalnumber representable in two different ways as a sum of two cubes: 1729 = 13 + 123 = 93 + 103. Generalizationsof this idea have spawned the notion of “taxicab numbers”. A very interesting article on another generalization iswritten by Henk Visser. The first part of this is reproduced on pp. 108-115 of this issue (the second part willappear in the December issue of the newsletter).

And now for something completely different. Although the BNAIC 2010 is finished meanwhile, we defer reportson it to the December issue (for space reasons), except for a report on the award of the SKBS-Strukton prize andthe minutes of the General Assembly. In this meeting it was announced that, in order to reduce costs, thenewsletter will undergo some changes, to be effectuated in 2011. Probably the most important one is that thefrequency of the newsletter will go back from six to four issues a year. To oppose the inherent drawback thatannouncements of BNVKI or other relevant events will be untimely, an accompanying up-to-date website willbe set up. Further, we also plan to reform the contents of the newsletter. For this, we are very interested in youropinion. What are the sections you (don’t) like? What are the types of contributions you miss? What othersuggestions for changes do you have?

Please send all suggestions to your editor-in-chief ([email protected]). They will be discussed inthe board and hopefully will lead to an improved newsletter, which is your newsletter after all. I’m lookingforward! For the 1729th serious suggestion, an appropriate gift will be arranged.

Wiki’s page on Ramanujan: http://en.wikipedia.org/wiki/Srinivasa_RamanujanHenk Visser’s home page: http://www.henkxvisser.nl/

mailto:[email protected]

http://en.wikipedia.org/wiki/Srinivasa_Ramanujan

http://www.henkxvisser.nl/


TABLE OF CONTENTS

1729 ........................................................................................................................................................... 106

Table of Contents ........................................................................................................................................ 107

BNVKI-Board News (Antal van den Bosch) ................................................................................................ 108

Ramanujaniana (part 1) (Henk Visser) ......................................................................................................... 108

The Magic of Monte-Carlo Tree Search (Mark Winands) ............................................................................ 115

Minutes of the BNVKI/AIABN General Assembly (Ann Nowé) .................................................................. 116

The 2010 SKBS-Strukton Prize (Jaap van den Herik) .................................................................................. 118

Ph.D. Thesis Abstracts ................................................................................................................................ 120Correctness of Services and their Composition (Niels Lohmann) .................................................... 120Integrative Modeling of Emotions in Virtual Agents (Ghazanfar Farooq Siddiqui) .......................... 121Ontology Enrichment from Heterogeneous Sources on the Web (Victor de Boer) ........................... 122Monte-Carlo Tree Search (Guillaume Chaslot) .............................................................................. 125Converting and Integrating Vocabularies for the Semantic Web (Mark van Assem) ........................ 128A Computational Approach to Content-Based Retrieval of Folk Song Melodies (Peter van Kranenburg).......................................................................................................... 130Allograph Based Writer Identification, Handwriting Analysis and Character Recognition (Ralph Niels) ....................................................................................................................... 132Ecological Automation Design, Extending Work Domain Analysis (Matthijs Amelink) .................. 132Context-Based Sound Event Recognition (Maria Niessen) ............................................................. 134

The Value of a Thesis (Jaap van den Herik) ................................................................................................. 136

SIKS (Richard Starmans) ............................................................................................................................ 138SIKS Basic Course “Research Methods and Methodology for IKS” ............................................... 138SIKS Basic Courses “Mathematical Methods for IKS” and “Knowledge Modeling” ....................... 139

Advertisements in the BNVKI Newsletter ................................................................................................... 139

Contact Addresses Board Members / Editors BNVKI Newsletter / How to Subscribe? / Submissions............ 140

Front cover: image of Srinivasa Iyengar Ramanujan, taken around 1900, by an unidentified author.

The deadline for the next issue is: December 1, 2010.


BNVKI-Board News

Antal van den Bosch

I think it is safe to say that BNAIC has reached aplateau: the Kirchberg Plateau, that is. BNAIC-2010 was held in the amazingly futuristic Kirchbergquarter of Luxembourg City, where the EuropeanUnion has some of its major bodies sit in high risingbuildings, and where the Luxembourg bankingworld has decided to cluster. Moreover, in theMusée d’Art Moderne Grand-Duc Jean (MudamLuxembourg) you can find some pretty alluring andthough provoking modern art, and we were fortu-nate to see the dark black cube of the Melia Hotelon the inside when we had the conference dinner.

BNAIC-2010 was great – it had all the ingredients,two wonderful invited speakers, a greatsurroundings, and a very lively atmosphere. Theboard wishes to heartily thank the organizers, agroup of people from various departments of theUniversity of Luxembourg, headed by Leon van derTorre, Pascal Bouvry, Eric Dubois, and ThibaudLatour, helped by an active troupe of helping hands.And as much as we loved being in the smallest ofour three Benelux countries, we have our headsturned now to BNAIC-2011, which in the GeneralAssembly was determined to be Ghent, Belgium.

In the General Assembly, the board also said good-bye to board members Sien Moens and RichardBooth. Sien and Richard, thank you both fordonating time and effort to our association; Sien forthe past five years, and Richard for an intense yearin which he was the liaison between the board andthe BNAIC-2010 organization. Thank you, both.

RamanujanianaPart 1

Henk VisserHaarlem

(It is the last working day of the year. Math, Comp,and Log are drinking coffee in the canteen of theL.E.J. Brouwer Institute, when Comp’s attention isdrawn by an interview in a newspaper in which thewell-known anecdote of Ramanujan is mentioned.Ramanujan called the number of Hardy’s taxicab,1729, extraordinary, because it is the smallestnumber that can be twice written as the sum of twocubic numbers.)

COMP. You are also a number freak, Math. Do youalso have such interesting insights?

MATH. I’m not so strongly impressed aboutRamanujan’s remark to Hardy as most people.After all, someone who knows the first fifteencubic numbers by heart, as Ramanujanundoubtedly did, can’t have failed to see that1000 + 729 is just 1 more than 1728.Nevertheless, I asked myself if I could invent asimilar, but less easily solvable problem. Myprinciple of variation helped me, as youunderstand. To present it in such a way that thevariation comes out clearly: instead of two sumsof powers of degree three, I took three sums ofpowers of degree two.

COMP. So you asked for the smallest numberthat can be written three times as the sum of twosquares?

MATH. Right, and because I know the first fortysquares by heart, I trusted that I could solve thisproblem.

LOG. But how could you be sure that there areany solutions at all?

COMP. What? That’s not a problem to me; Iknow the first four irreducible Pythagoreantriples by heart, and already the first three give asolution.

LOG. Show it!

COMP. We have 32 + 42 = 52, 52 + 122 = 132, and72 + 242 = 252, and the least common multiple of5, 13, and 25 gives a solution: 3252. I will writeit down for you (he takes a paper napkin):

(5·13·3)2 + (5·13·4)2 = (5·13·5)2

(5·5·5)2 + (5·5·12)2 = (5·5·13)2

(13·7)2 + (13·24)2 = (13·25)2

LOG. I am convinced!

MATH. 3252 is surely not the smallest numberwith the desired property, but nevertheless it isvery special. Comp, I thank you very much forthis outcome, because the smallest number that isthree times the sum of two squares is, believe itor not, 325!

LOG. I don’t believe in mathematical magic, butI admit that it is remarkable!

COMP. Please give the three sums, Math.

MATH. 325 = 324 + 1, 325 = 289 + 36, and 325= 225 + 100.

LOG. How did you find it?


MATH. The first sum I tried was 365, just for fun,because the year is almost over, and I saw or knewthat it is both the sum of 169 and 196, and the sumof 361 and 4, but there was no more to it. Then Itried 325, because it is 324 + 1, similar toRamanujan’s 1728 + 1, and I found the three sums.Thereafter I tried numbers smaller than 325, but thisyielded at most only numbers with two sums.

COMP. I will not underestimate your arithmeticalcapacities, Math, but still I will check yourconclusion on my computer if you don’t mind …

MATH. Good to hear that you keep working in thevacation period! I also have problems that botherme, so have a good time and see you next year!

LOG. Are you perhaps trying to invent moreRamanujanean problems, Math, or is there stillmore to tell about 325?

MATH. Yes and yes, and I can already answer yoursecond question: 325 is the triangular of a squarenumber, so it can be written as the sum of thesquares of two successive triangular numbers, justas the famous or notorious number 666.

LOG. I see (she writes on the napkin):

325 = 152 + 102, or t(s(5)) = s(t(5)) + s(t(4))

MATH. This can be represented in a perspicuousfigure, as you both know. It becomes my season’sgreetings picture (he takes a card out of his pocket):

Best wishes and a happy new year!

(They all go their own way.)

(When Math is still in his room, trying to bringsome order in his papers before leaving theInstitute, Comp enters.)

COMP. I checked your solution, and you wereright, 325 is the smallest number with the desiredproperty! But I did more, I looked for the smallest

number that is three times the sum of threesquares. It appeared to be 101, because 101 = 1 +36 + 64, 4 + 16 + 81, and 16 + 36 + 49.However, there are so many numbers which arethe sum of three squares, 110, 126, 134, 146, andso on, that I wondered what would be the use ofthis exercise.

MATH. Would it perhaps be better to follow myprinciple of variation, and look for the smallestnumber that is four times the sum of threesquares?

COMP. I will do my best, please stay in yourroom for a while! (He leaves, and returns half anhour later.)COMP. The smallest number that is four timesthe sum of three squares is 161, and the smallestnumber that is five times the sum of four squaresis 126. I don’t think it’s worthwhile to pursuethis further. Therefore I wrote a program inwhich the principle of generalization is appliedto your original problem: to find the smallestnumbers that are n times the sum of two squares.However, it isn’t guaranteed that the outcomeswill be optimal. I hereby give you the results inthe form I got them. I’m not sure about 11, 17,19, and 21. Therefore I omitted them. The othermissing values gave as yet no solutions. If thereare such, they are too large for my program.Moreover it got stuck after 32.

n = 15 = (s(1)+s(2))

n = 265 = (s(1)+s(8)) = (s(4)+s(7))

n = 3325 = (s(1)+s(18)) = (s(6)+s(17)) = (s(10)+s(15))

n = 41105 = (s(4)+s(33)) = (s(9)+s(32)) =(s(12)+s(31)) = (s(23)+s(24))

n = 58125 = (s(5)+s(90)) = (s(27)+s(86)) =(s(30)+s(85)) = (s(50)+s(75)) = (s(58)+s(69))

n = 65525 = (s(7)+s(74)) = (s(14)+s(73)) =(s(22)+s(71)) = (s(25)+s(70)) = (s(41)+s(62)) =(s(50)+s(55))

n = 7105625 = (s(36)+s(323)) = (s(80)+s(315)) =(s(91)+s(312)) = (s(125)+s(300)) =(s(165)+s(280)) = (s(195)+s(260)) =(s(204)+s(253))


n = 827625 = (s(20)+s(165)) = (s(27)+s(164)) =(s(45)+s(160)) = (s(60)+s(155)) = (s(83)+s(144)) =(s(88)+s(141)) = (s(101)+s(132)) = (s(115)+s(120))

n = 971825 = (s(1)+s(268)) = (s(40)+s(265)) =(s(65)+s(260)) = (s(76)+s(257)) = (s(104)+s(247)) =(s(127)+s(236)) = (s(160)+s(215)) =(s(169)+s(208)) = (s(188)+s(191))

n = 10138125 = (s(22)+s(371)) = (s(35)+s(370)) =(s(70)+s(365)) = (s(110)+s(355)) = (s(125)+s(350)) =(s(163)+s(334)) = (s(194)+s(317)) =(s(205)+s(310)) = (s(218)+s(301)) =(s(250)+s(275))

n = 12160225 = (s(15)+s(400)) = (s(32)+s(399)) =(s(76)+s(393)) = (s(81)+s(392)) = (s(113)+s(384)) =(s(140)+s(375)) = (s(175)+s(360)) =(s(183)+s(356)) = (s(216)+s(337)) =(s(228)+s(329)) = (s(252)+s(311)) =(s(265)+s(300))

n = 131221025 = (s(47)+s(1104)) = (s(105)+s(1100)) =(s(169)+s(1092)) = (s(264)+s(1073)) =(s(272)+s(1071)) = (s(425)+s(1020)) =(s(468)+s(1001)) = (s(520)+s(975)) =(s(561)+s(952)) = (s(576)+s(943)) =(s(663)+s(884)) = (s(700)+s(855)) =(s(744)+s(817))

n = 143453125 = (s(31)+s(1858)) = (s(110)+s(1855)) =(s(175)+s(1850)) = (s(350)+s(1825)) =(s(550)+s(1775)) = (s(625)+s(1750)) =(s(686)+s(1727)) = (s(815)+s(1670)) =(s(847)+s(1654)) = (s(970)+s(1585)) =(s(1025)+s(1550)) = (s(1090)+s(1505)) =(s(1201)+s(1418)) = (s(1250)+s(1375))

n = 151795625 = (s(5)+s(1340)) = (s(52)+s(1339)) =(s(179)+s(1328)) = (s(200)+s(1325)) =(s(325)+s(1300)) = (s(380)+s(1285)) =(s(467)+s(1256)) = (s(520)+s(1235)) =(s(563)+s(1216)) = (s(635)+s(1180)) =(s(676)+s(1157)) = (s(800)+s(1075)) =(s(808)+s(1069)) = (s(845)+s(1040)) =(s(940)+s(955))

n = 16801125 = (s(10)+s(895)) = (s(95)+s(890)) =(s(127)+s(886)) = (s(158)+s(881)) =(s(193)+s(874)) = (s(230)+s(865)) =

(s(241)+s(862)) = (s(335)+s(830)) =(s(370)+s(815)) = (s(430)+s(785)) =(s(458)+s(769)) = (s(463)+s(766)) =(s(529)+s(722)) = (s(545)+s(710)) =(s(554)+s(703)) = (s(610)+s(655))

n = 182082925 = (s(26)+s(1443)) = (s(134)+s(1437)) =(s(163)+s(1434)) = (s(195)+s(1430)) =(s(330)+s(1405)) = (s(370)+s(1395)) =(s(429)+s(1378)) = (s(531)+s(1342)) =(s(541)+s(1338)) = (s(558)+s(1331)) =(s(579)+s(1322)) = (s(702)+s(1261)) =(s(730)+s(1245)) = (s(755)+s(1230)) =(s(845)+s(1170)) = (s(894)+s(1133)) =(s(926)+s(1107)) = (s(1014)+s(1027))

n = 204005625 = (s(75)+s(2000)) = (s(147)+s(1996)) =(s(160)+s(1995)) = (s(336)+s(1973)) =(s(380)+s(1965)) = (s(405)+s(1960)) =(s(488)+s(1941)) = (s(565)+s(1920)) =(s(632)+s(1899)) = (s(700)+s(1875)) =(s(852)+s(1811)) = (s(875)+s(1800)) =(s(915)+s(1780)) = (s(1069)+s(1692)) =(s(1080)+s(1685)) = (s(1140)+s(1645)) =(s(1197)+ s(1604)) = (s(1260)+s(1555)) =(s(1325)+s(1500)) = (s(1344)+s(1483))

n = 2230525625 = (s(235)+s(5520)) =(s(525)+s(5500)) = (s(612)+s(5491)) =(s(845)+s(5460)) = (s(1036)+s(5427)) =(s(1131)+s(5408)) = (s(1320)+s(5365)) =(s(1360)+s(5355)) = (s(1547)+s(5304)) =(s(2044)+s(5133)) = (s(2125)+s(5100)) =(s(2163)+s(5084)) = (s(2340)+s(5005)) =(s(2600)+s(4875)) = (s(2805)+s(4760)) =(s(2880)+s(4715)) = (s(3124)+s(4557)) =(s(3315)+s(4420)) = (s(3468)+s(4301)) =(s(3500)+s(4275)) = (s(3720)+s(4085)) =(s(3861)+s(3952))

n = 245928325 = (s(63)+s(2434)) = (s(94)+s(2433)) =(s(207)+s(2426)) = (s(294)+s(2417)) =(s(310)+s(2415)) = (s(465)+s(2390)) =(s(490)+s(2385)) = (s(591)+s(2362)) =(s(690)+s(2335)) = (s(742)+s(2319)) =(s(849)+s(2282)) = (s(878)+s(2271)) =(s(959)+s(2238)) = (s(1039)+s(2202)) =(s(1062)+s(2191)) = (s(1201)+s(2118)) =(s(1215)+s(2110)) = (s(1290)+s(2065)) =(s(1410)+s(1985)) = (s(1454)+s(1953)) =(s(1535)+s(1890)) = (s(1614)+s(1823)) =(s(1633)+s(1806)) = (s(1697)+s(1746))


n = 2735409725 = (s(85)+s(5950)) = (s(338)+s(5941)) =(s(650)+s(5915)) = (s(782)+s(5899)) =(s(826)+s(5893)) = (s(901)+s(5882)) =(s(994)+s(5867)) = (s(1339)+s(5798)) =(s(1547)+s(5746)) = (s(1675)+s(5710)) =(s(1790)+s(5675)) = (s(1973)+s(5614)) =(s(2086)+s(5573)) = (s(2210)+s(5525)) =(s(2443)+s(5426)) = (s(2597)+s(5354)) =(s(2725)+s(5290)) = (s(2875)+s(5210)) =(s(3029)+s(5122)) = (s(3094)+s(5083)) =(s(3466)+s(4837)) = (s(3502)+s(4811)) =(s(3563)+s(4766)) = (s(3638)+s(4709)) =(s(3835)+s(4550)) = (s(4069)+s(4342)) =(s(4165)+s(4250))

n = 3229641625 = (s(67)+s(5444)) = (s(124)+s(5443)) =(s(284)+s(5437)) = (s(320)+s(5435)) =(s(515)+s(5420)) = (s(584)+s(5413)) =(s(835)+s(5380)) = (s(955)+s(5360)) =(s(1180)+s(5315)) = (s(1405)+s(5260)) =(s(1460)+s(5245)) = (s(1648)+s(5189)) =(s(1795)+s(5140)) = (s(1829)+s(5128)) =(s(1979)+s(5072)) = (s(2012)+s(5059)) =(s(2032)+s(5051)) = (s(2245)+s(4960)) =(s(2308)+s(4931)) = (s(2452)+s(4861)) =(s(2560)+s(4805)) = (s(2621)+s(4772)) =(s(2840)+s(4645)) = (s(3005)+s(4540)) =(s(3035)+s(4520)) = (s(3320)+s(4315)) =(s(3365)+s(4280)) = (s(3517)+s(4156)) =(s(3544)+s(4133)) = (s(3664)+s(4027)) =(s(3715)+s(3980)) = (s(3803)+s(3896))

MATH. Marvellous! And now it becomesinteresting: can we discover a regularity? The firstten outcomes suggest that we can divide allnumbers by 13, or, better, by 65. Moreover, do wehave to distinguish between even and odd numbers?This becomes my task for the vacation period!Exciting!

COMP. Have a good time, Math! ( He leaves theroom with a shrug of the shoulders.)

LOG. (Enters the room a few minutes later.) Stillworking, Math?

MATH. Log gave me a tremendous amount ofcomputations, and I am thinking them over.

LOG. Can I see them?

MATH. Here are the first ten items of Comp’s listwith the smallest numbers that are n times the sumof two squares. My outcome 325 is on this list:

n = 15 = (s(1)+s(2))

n = 265 = (s(1)+s(8)) = (s(4)+s(7))

n = 3325 = (s(1)+s(18)) = (s(6)+s(17)) = (s(10)+s(15))

n = 41105 = (s(4)+s(33)) = (s(9)+s(32)) =(s(12)+s(31)) = (s(23)+s(24))

n = 58125 = (s(5)+s(90)) = (s(27)+s(86)) =s(30)+s(85)) = (s(50)+s(75)) = (s(58)+s(69))

n = 65525 = (s(7)+s(74)) = (s(14)+s(73)) =(s(22)+s(71)) = (s(25)+s(70)) = (s(41)+s(62)) =(s(50)+s(55))

n = 7105625 = (s(36)+s(323)) = (s(80)+s(315)) =(s(91)+s(312)) = (s(125)+s(300)) =(s(165)+s(280)) = (s(195)+s(260)) =(s(204)+s(253))

n = 827625 = (s(20)+s(165)) = (s(27)+s(164)) =(s(45)+s(160)) = (s(60)+s(155)) = (s(83)+s(144)) =(s(88)+s(141)) = (s(101)+s(132)) =(s(115)+s(120))

n = 971825 = (s(1)+s(268)) = (s(40)+s(265)) =(s(65)+s(260)) = (s(76)+s(257)) =(s(104)+s(247)) = (s(127)+s(236)) =(s(160)+s(215)) = (s(169)+s(208)) =(s(188)+s(191))

n = 10138125 = (s(22)+s(371)) = (s(35)+s(370)) =(s(70)+s(365)) = (s(110)+s(355)) =(s(125)+s(350)) = (s(163)+s(334)) =(s(194)+s(317)) = (s(205)+s(310)) =(s(218)+s(301)) = (s(250)+s(275))

LOG. It’s curious that there are no doubles. Infact, the smallest number that is the sum of twosquares is not 5, but 2. For n = 2, I seeimmediately 50, because 50 = 49 + 1, and 50 =25 + 25. Comp’s solution gives the smallestnumbers that are n times the sum of two differentsquares! I hope that it doesn’t affect yoursolution for n = 3, Math!

MATH. (After some scribbling on theblackboard.) No, that’s all right. Nevertheless, I


will stay with Comp’s outcomes, because they seemto have some nice properties. I factorized allreliable numbers and this resulted in the followingtable:

n = 15 = 5

n = 265 = 5·13

n = 3325 = 52·13

n = 41105 = 5·13·17

n = 58125 = 54·13

n = 65525 = 52·13·17

n = 7105625 = 54·132

n = 827625 = 53·13·17

n = 971825 = 52·132·17

n = 10138125 = 54·13·17

n = 12160225 = 52·13·17·29

n = 131221025 = 52·132·172

n = 143453125 = 56·13·17

n = 151795625 = 54·132·17

n = 16801125 = 53.13.17.29

n = 182082925 = 52·132·17·29

n = 204005625 = 54·13·17·29

n = 2230525625 = 54·132·172

n = 245928325 = 52·13·17·29·37

MATH. Wait, an email message from Comp.(Math reads it.)

LOG. What does Comp write?

MATH. He found more numbers, which healready factorized. I’ll write them down, togetherwith the values that we already have:

n = 2752·132·172·29

n = 2856·13·17·29

n = 3054·132·17·29

n = 3253·13·17·29·37

n = 3652·132·17·29·37

n = 4054·13·17·29·37

n = 4554·132·172·29

n = 4852·13·17·29·37·41

And, last but not least, surprisingly:

n = 11510·13

LOG. The powers of 5 are never less than thoseof the other prime numbers of the form 4k + 1.Not all combinations occur. Do you have anexplanation, Math?

MATH. I presume that the combinations whichare skipped, belong to numbers with one of therequired properties, without being minimal ofcourse.

LOG. Let us try 52·132. How do we find thedesired representations?

MATH. Believe it or not, the answer can befound on Internet, by what is called thecomputational knowledge machineWolframalpha. (Math goes to the site and hefinds the answer.) Look:


LOG. That’s nice, and now 53·13.

MATH. Do you need more examples?

LOG. Perhaps 53·132 and 53·133?

MATH. (He fills in 21125 and 274625.)

LOG. Again a message from Comp. (She opens it.)Interesting! Comp had a brilliant insight! He foundregularities by ordering the numbers in a specialway, first 5, 5·17, 5·17·29, 5·17·29·37,5·17·29·37·41, then 52, 52·17, 52·17·29, 52·17·29·37,and so on, each time taking the next power of 5 andomitting the factor 13. He encoded the prime factorsby their exponents (with the numbers behind theminus signs giving the number of representations assums of two different squares), and got thefollowing results (She prints them out.)

5: 1 - 185: 101 - 22465: 1011 - 491205: 10111 - 83739405: 101111 - 16

25: 2 - 1425: 201 - 312325: 2011 - 6456025: 20111 - 1218697025: 201111 - 24

125: 3 - 22125: 301 - 461625: 3011 - 82280125: 30111 - 1693485125: 301111 - 32

625: 4 - 210625: 401 - 5308125: 4011 - 1011400625: 40111 - 20467425625: 401111 - 40

3125: 5 - 353125: 501 - 61540625: 5011 - 1257003125: 50111 - 24

15625: 6 - 3265625: 601 - 77703125: 6011 - 14285015625: 60111 - 28

MATH. It seems that the bare powers of 5 mustbe omitted in order to get complete regularities.

85: 101 - 22465: 1011 - 491205: 10111 - 83739405: 101111 - 16

425: 201 - 312325: 2011 - 6456025: 20111 - 1218697025: 201111 - 24

2125: 301 - 461625: 3011 - 82280125: 30111 - 1693485125: 301111 - 32

10625: 401 - 5308125: 4011 - 1011400625: 40111 - 20467425625: 401111 - 40

53125: 501 - 61540625: 5011 - 1257003125: 50111 - 24

265625: 601 - 77703125: 6011 - 14285015625: 60111 - 28

LOG. Comp missed two outcomes, namely57003125·41 and 285015625·41. This seems toprovide an opportunity for checking the rule thatsubjoining the next prime factor of the form 4k +1 leads to a doubling of the number of sums.

MATH. (He does a little computing and consultsWolframalpha.)

57003125·41 = 2337128125

Promising!

285015625·41 = 11685640625


(???)

That’s a pity, Wolfram gives no representations,although it mentions that 11685640625 is thehypotenuse of 16 primitive Pythagorean triples, butthat’s another topic. Perhaps the predicted numberof 56 representations is too large for thecomputational knowledge machine? Still, the 48representations of 55·17·29·37·41 seem to givesufficient evidence for trusting the doubling rule.We must thank Comp!1 (He writes a reply to Compin which he congratulates him with what heachieved.)

LOG. Nevertheless, I want a proof!

MATH. I will think it over ... and we have not evensolved our original problem of finding regularitiesin the minimal numbers ... Perhaps a comparisonbetween their table and Comp’s last one will help.

LOG. It strikes me that the minimal numbers allhave the factor 13, whereas this factor is completelyabsent in the regularities.

MATH. Comp arranged the numbers according tothe powers of the factor 5 and the increasing seriesof factors of the form 4k + 1. It seems that we mustdo the same with the minimals, to begin with. Firstthe exponent 2, starting with the value for n = 3, myfamous 325. I will write it as m(3) = 325. Then weget:

m(3) = 52·13m(6) = 52·13·17m(12) = 52·13·17·29m(24) = 52·13·17·29·37m(48) = 52·13·17·29·37·41

LOG. I understand how it continues with 53:

m(8) = 53·13·17m(16) = 53·13·17·29m(32) = 53·13·17·29·37

MATH. Isn’t it simple? I continue:

m(5) = 54·13m(10) = 54·13·17m(20) = 54·13·17·29m(40) = 54·13·17·29·37

LOG. Yes, and finally:

1 All credit to Dr. Jeroen Donkers, who not onlywrote the computer programs, but also discoveredthe surprising theorem.

m(14) = 56·13·17m(28) = 56·13·17·29

MATH. Let’s now try the cases in which thefactor 13 has the exponent 2, to begin with n = 9:

m(9) = 52·132·17m(18) = 52·132·17·29m(36) = 52·132·17·29·37

LOG. Then we also have:

m(15) = 54·132·17m(30) = 54·132·17·29

MATH. So far, so good, but we skipped somevalues, not only 1, 2, and 4, but also 7, 13, and22, not to mention 11. We can investigate whathappens with them, consulting our friendWolfram! The values of 1, 2, and 4 are 5, 5·13,and 5·13·17. Let’s see what happens with5·13·17·29, that is, 32045.

LOG. That’s OK, and wholly in the spirit ofComp! But 32045 isn’t minimal, that was 27625.It’s confirmed by Wolfram:

MATH. m(7) = 54·132, but I don’t know what todo with it. The next values are already givenaway, 54·132·17 = m(15), and 54·132·172 = m(22).It’s true that we can go back from m(14) =56·13·17 and m(28) = 56·13·17·29, to get 56·13,and verify that it has 7 representations as a sumof 2 squares, but 56·13 is clearly larger than54·132.

LOG. We know that 510·13 has 11representations, but is it minimal? Moreover, therelation with m(22) = 54·132·172 isn’t clear. I’mafraid that we do not get any further. So far theresults are impressive, but they leave much to bedesired. And now I need a drink.

MATH. Wait a minute! The beautiful value for11 gives me the idea that we should investigateall preceding numbers of the form 5k·13 with thehelp of Wolfram, in so far as we don’t alreadyhave their number of representations . I use theletter n in stead of m, because I expect that somevalues are not minimal.


(By consulting Wolfram when necessary, Mathcomes to the following table.)

n(50·13) = 1n(51·13) = 2n(52·13) = 3n(53·13) = 4n(54·13) = 5n(55·13) = 6n(56·13) = 7n(57·13) = 8n(58·13) = 9n(59·13) = 10n(510·13) = 11

MATH. Now we see how Comp’s solution foreleven representations is generated!

LOG. Amazing! It gives every number ofrepresentations! Maybe we can try to prove thegeneral theorem, before going to the otherregularities?

MATH. Seems plausible, but let’s first take a drop.

(He opens a cupboard and makes preparations fora drink.)

LOG. If we stay in your room, we can investigatethe n(5k·132)’s, the n(5k·133)’s, the n(5k·134)’s, andsee how far we get on with them.

MATH. (filling the glasses) Go ahead!

LOG. (a couple of minutes later, helped byWolfram) Here you are:

n(50·132) = 1n(51·132) = 3

n(52·132) = 4n(53·132) = 6

n(54·132) = 7n(55·132) = 9

n(56·132) = 10n(57·132) = 12

n(58·132) = 13

n(50·133) = 2n(51·133) = 4n(52·133) = 6n(53·133) = 8n(54·133) = 10n(55·133) = 12

n(50·134) = 2

n(51·134) = 5n(52·134) = 7

n(53·134) = 10n(54·134) = 12

n(55·134) = 15n(56·134) = 17

MATH. 17 representations in seven steps,incredible! This deserves a toast!

(After quite a number of drinks, they end thesession in high spirits, and go home.)

[ To be continued ]

The Magic ofMonte-Carlo Tree Search

September 29, 2010

Mark WinandsDKE, Maastricht University

On September 29, in collaboration with SIKSand the BNVKI, the Department of KnowledgeEngineering (DKE) at Maastricht Universityorganized a symposium called “The Magic ofMonte-Carlo Tree Search”. The symposium washeld in honour of Guillaume Chaslot’s Ph.D.defence. The topic of his thesis (and therefore ofthe symposium as well) was Monte-Carlo TreeSearch (MCTS). Approximately 20 participantsattended the five talks. An outline of the fivetalks is given below.

YNGVI BJÖRNSSON (REYKJAVIK UNIVERSITY)Yngvi Björnsson gave an inspiring talk abouteffectively using MCTS in General Game-Playing (GGP) programs. The aim of GGP is tocreate intelligent agents that can automaticallylearn how to play many different games at anexpert level without any human intervention.One of the main challenges such agents face is toautomatically learn knowledge-based heuristicsin real-time, whether for evaluating gamepositions or for search guidance. In recent years,GGP agents that use MCTS simulations toreason about their actions have becomeincreasingly more popular. One nice property ofGGP as a testbed for research purposes is thatone can use such game-playing programs to


study and contrast the effectiveness of MCTS onmany different types of games. Also of an interestis that Go and GGP game playing programs,although both prominent examples of an effectiveuse of MCTS, exhibit widely differentcharacteristics. In computer Go, before coming upwith a move decision, programs run a high numberof simulations (in the hundreds of thousands) thatare controlled using pre-defined and (typically)hand-crafted search-control knowledge.Conversely, in GGP, where state-spacemanipulation is more computing intensive, movedecision must be based on much fewer simulations(few thousands), as well as the required search-control knowledge must be learned online duringplaying games.

PIM NIJSSEN (MAASTRICHT UNIVERSITY)MCTS is becoming increasingly popular for playingmulti-player games. In his talk, Pim Nijssenproposed two enhancements for MCTS in multi-player games. They are called Progressive Historyand Multi-Player Monte-Carlo Tree Search Solver(MP-MCTS-Solver). He analyzed the performanceof these enhancements in two different multi-playergames: Focus and Chinese Checkers. Based onexperimental results he showed that ProgressiveHistory is a considerable improvement in bothgames. MP-MCTS-Solver, using the standardupdate rule, is a genuine improvement in Focus.

HENDRIK BAIER (OSNABRÜCK UNIVERSITY)Hendrik Baier discussed a new simulation strategyfor MCTS. The MCTS algorithm builds a searchtree by playing many simulated games (playouts).Each playout consists of a sequence of moveswithin the tree followed by many moves beyond thetree. Moves beyond the tree are generated by abiased random simulation strategy. The recentlypublished last-good-reply strategy of Peter Drakeselects moves that, in previous playouts, have beensuccessful replies to immediately preceding moves.In his talk, Hendrik Baier presented a modificationof this strategy that not only remembers moves thatrecently succeeded, but also immediately forgetsmoves that recently failed. This modificationprovides a large improvement in playing strength.He also showed that responding to the previous twomoves is superior to responding to the previous onemove. Remembering the win rate of every replyperforms much worse than simply remembering thelast good reply (and worse than not storing goodreplies at all).

BRUNO BOUZY (UNIVERSITÉ PARIS DESCARTES)Bruno Bouzy presented a new heuristic searchalgorithm based on mean values for real-timeplanning, called Mean-based Heuristic Search for

real-time Planning (MHSP). It combines theprinciples of MCTS with heuristic search inorder to create a real-time planner. In contrast toMCTS, at leaf nodes MHSP replaces thesimulations (playouts) by heuristic values givenby graph-planning techniques. When theheuristic is admissible, the initial mean values ofnodes are optimistic, which is a correct way ofguiding exploration. In his talk Bruno Bouzyshowed promising results of the application ofMHSP to the following planning domains:blocksworld, ferry, gripper, and satellite.

GUILLAUME CHASLOT (GOOGLE)Guillaume Chaslot answered in his talk thequestion how to enhance MCTS in such a waythat a tournament program improves itsperformance in Go. First, he improved theknowledge in the simulation strategy by learningfrom move evaluations. Second, he enhanced theselection strategy by proposing progressivestrategies for incorporating knowledge. Third, heapplied the Cross-Entropy Method to optimizethe search parameters of an MCTS program insuch a way that its playing strength is increased.Fourth, he designed Meta-MCTS to generate anopening book that improves the performance ofan MCTS program.

All five presentations can be downloaded at:www.unimaas.nl/games/symposium2010/

Minutes of theBNVKI/AIABN General Assembly

Tuesday October 26, 2010Luxembourg

Ann Nowé

Present: Antal van den Bosch (Chair), VirginiaDignum, Jos Uiterwijk, Richard Booth, AnnNowé, and 24 members.

1. OpeningChair Antal van den Bosch opens the meeting at13:45.

2. Minutes of the BNVKI General Assemblyof October 31, 2009

The minutes are approved.

3. AnnouncementsAnn Nowé joined the ECCAI board and hastaken over the task of the former treasurerHendrik Blockeel.

http://www.unimaas.nl/games/symposium2010/


Leon van der Torre says that the people ofLuxembourg are positive about joining BNVKI. Heclarifies the role of the different partners involvedin the organisation of BNAIC 2010: the CRP whichis mainly responsible for the logistic support, whilethe university partners have taken care of thescientific issues. Leon recognises the sponsors.

The organisers received 76 submissions (36 A-papers, 34 B-papers and 6 demos), of which 67have been accepted (88.16%) (50 oral, 11 postersand 6 demos). On 22/10/10 the organisers received99 registrations (50 from The Netherlands, 18 fromBelgium, 22 from Luxembourg, and 9 others).

Leon van der Torre continues by explaining that thefunding by the Luxembourg science foundationimplies that BNAIC’10 cannot make profit, so theregistration fees were kept low. This yearproceedings are available on a USB stick, printedproceedings can be optionally ordered. The boardbelieves that providing electronic proceedings is theway to go, it also is the tendency at big conferences.

Some members suggest that for next editions papersshould be made available one week ahead, sopeople can make printouts, and a booklet withabstracts is also advisable. This idea will beconsidered for the next BNAIC.

4. Financial report 2009Virginia Dignum, treasurer, reports on the financialsituation of the BNVKI. She explains that thedifference between what was estimated andrealised, is mainly due to increasing secretarialcosts (increase of prices and the fact that there are 2locations (Maastricht and Tilburg). Before, NWOwas supporting the BNVKI via Maastricht. Thisbudget was used to cover the secretarial costs. Now,NWO supports the BNVKI by sponsoring theregistration of Dutch students attending the BNAIC.

Although some money from the reserve was used,the assets are still quite large: 40,038 euro.However, the board recognises that the secretarialcosts should be reduced and will discuss how thiscan be done. One idea is that the number ofnewsletters will be reduced to 4 per year, and willbe complemented with a monthly updated agendaon the website.

The auditing committee consisting if Menno vanZaanen and Birna van Riemsdijk approved theexpenses.

5. Auditing committee 2010Treasurer Virginia Dignum proposes the newauditing committee, which checks the financialreport to be delivered at the next General Assembly,

to consist of Silja Renooij and Mark Hagedoorn.This proposal is accepted by the assembly.

6. Progress report 2010 and plans for 2011Antal van den Bosch explains that themembership of BNVKI implies a membership ofECCAI, which allows a reduction of theregistration fee at ECAI and free access to AICommunications.

In The Netherlands the NSVKI, the organisationof AI students, is again active and is publishingthe newsletter De Connectie and organising theBNAIS student conference. Annette ten Teije isresponsible within the board to keep the linkwith the students.

Antal van den Bosch provides an overview ofpast Benelux AI events.

• The Magic of Monte-Carlo Tree Search,Symposium, September 29, 2010,Department of Knowledge Engineering,Maastricht University, The Netherlands

• Confluences in Models of Rationality,Workshop, May 30, 2010, KatholiekeUniversiteit Leuven, Belgium

• BENELEARN 2010, the annual machine-learning conference of Belgium and TheNetherlands, May 27-28, 2010, KatholiekeUniversiteit Leuven, Belgium

• SOKS Symposium on the link betweenComputational Intelligence and the SemanticWeb, April 28-29, 2010, Vrije UniversiteitAmsterdam, The Netherlands

• Formal Models of Norm Change 2,Workshop, January 18-19, 2010, Universityof Amsterdam, The Netherlands

• ESAW 2009, The 10th Annual InternationalWorkshop “Engineering Societies in theAgents’ World”, November 18-20, 2009,Utrecht University, The Netherlands

• SRL-2009 (International Workshop onStatistical Relational Learning), ILP-2009(19th International Conference on InductiveLogic Programming), and MLG-2009 (7th

International Workshop on Mining andLearning with Graphs), July 2-4, 2009,Katholieke Universiteit Leuven, Belgium

• INTETAIN 09, The 3rd InternationalConference on Intelligent Technologies forInteractive Entertainment, June 22-24, 2009,Amsterdam, The Netherlands

• BENELEARN 09, The 18th Annual Belgian-Dutch Conference on Machine Learning,May 18-19, 2009, Tilburg University, TheNetherlands


• GALP, ICR International Symposium onGames, Argumentation, and LogicProgramming, April 23-24, 2009, Luxembourg

• Rich Cognitive Models for Policy Design andSimulation, January 12-16, 2009, LorentzCentre, Leiden, The Netherlands

Antal van den Bosch invites the audience topropose ideas of what the BNVKI can offer extra toits members, to attract also members who didn’tattend BNAIC.

Some suggestions from the audience were:

• The board could consider to support theorganisation of ECAI or similar events in theBenelux.

• The BNVKI should be represented at localevents such as SIREN.

• The BNVKI should be represented in socialnetworks such as LinkedIn, so a memberbenefits by a higher visibility.

• A list of ECCAI fellows should be available onthe BNVKI website.

Finally, a member remarks that the secretarial costsof 60% is high. The board agrees with this and hasalready ideas to reduce it and in the budget for nextyear this percentage is lowered.

7. BNVKI boardAntal van den Bosch gives an overview of thecurrent board members and informs the meetingthat Sien Moens is stepping out the board, becauseher term is finished. Also Richard Booth wants toresign; he was the first postdoc member who joinedthe board and the idea was that postdocs wouldtypically serve a shorter term. At the next generalassembly some new board members will be elected.

8. BNAIC’11The board is happy to announce that KatjaVerbeeck and Partick De Causmaeker of theCODeS group, KULeuven, have accepted toorganize BNAIC 2011, which will be held onNovember 3 and 4, 2011 at Ghent, Campus KaHoSt. Lieven. Due to lack of time, the presentation ispostponed till the closing session.

9. End of meetingThere are no more comments or questions. Themeeting is closed at 14:30.

The 2010 SKBS-Strukton Prize

Jaap van den HerikDirector of SKBS

The Foundation for Knowledge Based Systems(SKBS) and the Company Strukton continuedtheir policy of awarding the SKBS-Struktonprize to the best demonstration of the Demo-session of the BNAIC 2010. Since the BNAICwas held as far as in Luxembourg, the companyStrukton was unable to attend this BNAIC. Theactive BNVKI member Bas Obladen (who is aregular visitor since the mid 1990s) and CarlosF. Bosma MSc had to give preference to theirbusiness activities. Yet, their interest in newideas is so intensive that they confirmed tosupport the SKBS-Strukton prize also in 2010with an amount of Euro 350, bringing the fullprize to Euro 700. We are grateful for thisgenerous gesture and wish the company aprosperous climate for the months to come. Wehope to see Obladen and Bosma again in Ghentin Belgium on November 3 and 4, 2011 at the23rd BNAIC.

The 2010 referee committee consisted of Jaapvan den Herik (chair), Han La Poutré (formerchair BNVKI), Hendrik Blockeel (Belgium,Leuven), Leon van der Torre (Luxembourg),Birna van Riemsdijk (Delft University ofTechnology), and Niek Wijngaards(businessman and researcher at Thales). Jaap vanden Herik acted as non-voting chair owing to theinvolvement of Jeroen Janssens and Eric Postma.

The referee committee had to consider sixsubmissions which were eligible for the SKBS-Strukton prize. In Table 1 we list them by topic(in the order of their publication in theConference Program BNAIC 2010).

Since 1999 we have seen many differentappearances of the Demo-session. The commoncharacteristic is the emphasis on being “anindustrial exhibition”. Up to 2006 the prizemoney was provided by SKBS only. TheFoundation for Knowledge Based Systemsoriginates from the late 1980s as a foundationwithin SPIN (Stimulerings Projectteam InNederland). The Foundation SNN (StichtingNeurale Netwerken) is another well-knownmember of the former SPIN. They supportedSKBS financially with augmenting the SKBSprize in 2007. In 2008, the industrial partnerStrukton announced its willingness to participatein the prize funding. The extra contribution was


gratefully accepted. As stated above they continuedthis policy in 2009 and 2010.

1. Disaster Victim Identification SystemWillem Burgers, Wim Wiegerinck, and BertKappen2. Creating artificial vessel trajectories withPrestoJeroen Janssens, Hans Hiemstra, and EricPostma3. Road Network Hierarchy Generation byDistributed Agents for a RoutingApplicationJoris Maervoet, Lode Blomme, Katja Verbeeck,and Greet Vanden Berghe4. RoboCup.be – A testbed for intelligentsoccer strategiesTim Vermeulen and Katja Verbeeck5. ConfigNow: a knowledge based approachto configuration softwareHanne Vlaeminck, Kumar Abhinav, JoostVennekens, Marc Denecker, and Johan Wittocx6. The IDP SystemJohan Wittocx, Broes De Cat, and MarcDenecker

Table 1: The 2010 candidates of the SKBS-Strukton prize.

In 2010, six submissions were exhibited in thedemonstration room Salle Nancy where the BNAICtook place. All six were full-fledgeddemonstrations. It was really a pleasure to walkalong the demos and to discuss them with the standholders. Again, the quality and maturity ofpresentation, particularly the quality of ideas, hasgrown considerably over the last year.

The referee committee had a difficult task. Theprocedure went in shifts: from six we reduced thenumber of candidates to three and then to two andfinally we arrived at one.

The members of the referee committee were invitedto score on (a) the originality of the submission, (b)the applicability (scalability), (c) the AI content, (d)the actual applicability in society, (e) thepresentation, (f) the applicability in other domains,(g) the applicability by other persons, and (h) opensource.

There was a long discussion which finally led to thefollowing winners: Willem Burgers, WimWiegerinck, and Bert Kappen for their demoBonaparte Disaster Victim Identification System.The second place was for the IDP system, and thethird place for Presto.

Thus, the winning demo was BONAPARTE, whichhas been used by the NFI (Netherlands Forensic

Institute) for identification of the victims of theTripoli Air Crash in the beginning of 2010. Thenit was a -side program, but now it is accepted asan asset to help the NFI, Schiphol, and theKLPD with their daily work.

Congratulations to the Nijmegen team (RUN),especially to Bert Kappen (SNN) who has wonfor the second time the SKBS-Strukton prize(SKBS 2003, SKBS-Strukton 2010).

In Table 2 we provide an overview of thewinners of the SKBS-Strukton prize so far.

1999 MaastrichtM. van Wezel, J. Sprenger, R. van Stee, and H. LaPoutréNeural Vision 2.0 – Exploratory Data Analysis withNeural Networks

2000 Kaatsheuvel (shared prize)E. ZopfiHKTG. SchramLubeSelect

2001 AmsterdamAlexander Ypma, Rob Kleiman, Jan Valk, and BobDuinMINISOM – A System for Machine Health Monitoringwith Neural Networks

2002 LeuvenF. Brazier, D. Mobach, and B. OvereinderAgentScape Demonstration

2003 NijmegenBert Kappen, Wim Wiegerinck, Ender Akay, MarcelNijman, Jan Neijt, and André van BeekPromedas: A Diagnostic Decision Support System

2004 GroningenWouter TeepeThe Secret Prover: Proving Possession of ArbitraryFiles While not Giving Them Away

2005 BrusselsGerald de JongFluidiom: The Evolution of Locomotion

2006 NamurMarion Verduijn, Niels Peek, Peter Rosseel, Evert deJonge, and Bas de MolProcarsur: A System for Prognostic Reasoning inCardiac Surgery

2007 UtrechtTim Harbers, Rob van der Veen, and Marten den UylSentient Demonstration BNAIC 07: Vicavision


2008 Enschede (shared prize)Joris Maervoet, Patrick De Causmaecker, and GreetVanden BergheA Generic Rule Miner for Geographic DataandDennis Reidsma and Anton NijholtTemporal Interaction between an Artificial OrchestraConductor and Human Musicians

2009 EindhovenTom van Bergen, Maarten Brugmans, Bart Dohmen,and Niels MolenaarCobes: The clean, safe and hospitable metro

2010 LuxembourgDisaster Victim Identification SystemWillem Burgers, Wim Wiegerinck, and Bert Kappen

Table 2: Overview of SKBS-Strukton prize winners.

Correctness of Services and theirComposition

Ph.D. thesis abstractNiels Lohmann

Promotores: Prof.dr. W.M.P. van der Aalst, Prof.dr.K. Wolf

Copromotor: Dr. N. SidorovaDate of defense: September 27, 2010

Service-oriented computing (SOC) is an emergingparadigm of system design and aims at replacingcomplex monolithic systems by a composition ofinteracting systems, called services. A serviceencapsulates self-contained functionality and offersit over a well-defined, standardized interface.

This modularization may reduce both complexityand cost. At the same time, new challenges arisewith the distributed execution of services indynamic compositions. In particular, thecorrectness of a service composition depends notonly on the local correctness of eachparticipating service, but also on the correctinteraction between them. Unlike in a centralizedmonolithic system, services may change and arenot completely controlled by a single party.

We study correctness of services and theircomposition and investigate how the design ofcorrect service compositions can besystematically supported. We thereby focus onthe communication protocol of the service andapproach these questions using formal methodsand make contributions to three scenarios ofSOC.

The correctness of a service compositiondepends on the correctness of the participatingservices. To this end, we (1) study correctnesscriteria which can be expressed and checked withrespect to a single service. We validate servicesagainst behavioral specifications and verify theirsatisfaction in any possible service composition.In case a service is incorrect, we providediagnostic information to locate and fix the error.

In case every participating service of a servicecomposition is correct, their interaction can stillintroduce problems. We (2) automatically verifycorrectness of service compositions. We furthersupport the design phase of service compositionsand present algorithms to automatically completepartially specified compositions and to fixincorrect compositions.

A service composition can also be derived froma specification, called choreography. Achoreography globally specifies the observablebehavior of a composition. We (3) present analgorithm to deduce local service descriptionsfrom the choreography which – by design –conforms to the specification.

All results have been expressed in terms of aunifying formal model. This not only allows toformally prove correctness, but also makesresults independent of the specifics of concreteservice description languages. Furthermore, allpresented algorithms have been prototypicallyimplemented and validated in experiments basedon case studies involving industrial services.

PH.D. THESIS ABSTRACTS


Integrative Modeling of Emotions inVirtual Agents

Ph.D. thesis abstractGhazanfar Farooq Siddiqui

Promotor: Prof.dr. J. TreurCopromotores: Dr. T. Bosse, Dr. J.F. HoornDate of defense: September 28, 2010

In recent years, researchers have becomeincreasingly interested in the application ofintelligent virtual agents in various domains.Intelligent virtual agents (IVAs) are autonomous,graphically embodied agents in a virtualenvironment that are able to interact intelligentlywith the environment, other IVAs, or with humanusers. Recently, much research has been dedicatedto developing virtual agents with more realisticgraphical representations. However, the affectiveproperties of such agents are usually rather limited,and not very human-like. For example, althoughmany IVAs currently have the ability to somehowshow emotions by means of different facialexpressions, it is quite difficult for them to show theright emotion at the right moment. One step further,it is even more difficult for them to actuallyunderstand and react empathically to the emotionalstate of other agents. This is in conflict with therequirement of virtual agents to closely mimichuman affective behavior. Several studies in SocialSciences have shown that this is an importantprerequisite for an agent to increase humaninvolvement in a virtual environment. Therefore,existing systems based on IVAs are not as effectiveas they could be. Properties that they typically lackare the ability to show emotions (not only in termsof facial expression, but also in terms of behavior),

in relation to insight in each other’s and humans’cognitive and affective states.

To deal with this problem, some authors proposeto increase the affective properties of interactivesoftware agents by using knowledge frompsychology and cognitive science as a basis forcomputational modeling of the cognitive andaffective processes involved. Recently, a varietyof such computational models have beendeveloped for different aspects of humanbehavior. Examples include models forreasoning processes, visual attention, emotionregulation, mindreading, stress and workload,and moods. If such computational models areavailable in a formal format, this opens thepossibility to equip IVAs with them. However,the step from existing computational models(that are mostly used for simulation purposes) tomodels that can directly be plugged into a virtual(3D) environment in such a way that the IVAsbehave according to the cognitive model, is anontrivial one. In reality, this step involves aniterative process, consisting of, among others,the following sub-tasks: refinement of thecomputational model, translation to a specificprogramming environment, and testing andevaluation of the resulting model in the virtualsetting.

The main research goal of this thesis is toexplore how computational models of affect canbe integrated within virtual agents. To this end,different theories were taken from differentfields (e.g., Social Sciences, Psychology,Economics and Finance) and have beencombined to develop integrative models ofaffect. Most of these models have first been usedfor simulation, to test whether their overallbehavior was satisfactory. For this, variousmodeling environments have been used, such asLEADSTO and C++. Next, these models havebeen incorporated within applications related tohealth care, games and business context. Theseapplications have been developed usingJavaScript, in combination with the Vizardtoolkit. Finally, for some applications, user testshave been performed. The preliminary results ofthese tests show that through the developedmodels, users are more involved in theapplications.

In the first part of the thesis, the extent to whichan agent becomes involved with another agent,or stays at a distance from it, is modeled, byformalizing in a computational manner the(previously informally described) I-PEFiCmodel. These extents depend upon differentaspects of the other agent such as ethics (good or


bad), aesthetics (beautiful or ugly), epistemics orrealism (how realistic or unrealistic the other agentis), similarity (resemblances between the twoagents) and affordances the other agent offers as anaid or obstacle for task performance of the agent.Second, the integration of three models of affect isaddressed. The approach taken in this part is toselect three of the more influential models, whichshare that they can be used to enhance believabilityof virtual characters: CoMERG (ComputationalModel of Emotion Regulation based on Grosstheory), EMA (Emotion and Adaptation) and I-PEFiCADM (I-PEFiC extended with a module forAffective Decision Making). Then, we focus onmodeling affective states of a person in economicalcontext. This economical context is considered alarge-scale multi-agent system consisting ofthousands or millions of other agents. In this part itis described how a person’s involvement with theseother agents in the form of individual investmentdecisions depends on a personal risk profile, thestate of greed of the person, and the state of theworld economy. Finally, in the last part of thethesis, applications of virtual agents that showemotions in the domains of health care, business,and games are presented and evaluated.

Ontology Enrichment fromHeterogeneous Sources on the Web

Ph.D. thesis abstractVictor de Boer

Promotor: Prof.dr. B.J. WielingaCopromotor: Dr. M. van SomerenDate of defense: September 30, 2010

This thesis is about semi-automatic methods thatextract structured information from unstructureddocuments on the Web.

The Semantic Web is a proposed extension ofthe current World Wide Web (WWW). Wherethe WWW is a web of documents connectedthrough hyperlinks, the Semantic Web is a webof interlinked data. The documents on the WWWare designed to be read and interpreted byhumans. Since the information on the SemanticWeb appears in a formalized form, it can beinterpreted and reasoned with by computerapplications.

An example of the advantages of havinginformation available in this formalized form canbe seen in the MultimediaN E-Culturedemonstrator. This application uses backgroundknowledge to connect metadata descriptions ofobjects of Dutch cultural heritage institutions toeach other. The fact that the terms that are usedas metadata in the object descriptions areinterlinked allows the application to performsearch and browsing tasks that can not beperformed by existing text-based search systems.If, for instance, we know that Vincent van Goghis considered a Post-Impressionist, we candeduce that his paintings have some relation tothis style and to other works of other Post-Impressionists.

PROBLEM AND APPROACHThe information or knowledge used in theSemantic Web can be divided into twocategories: The first category is that of meta-knowledge that describes which classes exist in adomain and which relations hold between theseclasses. For the cultural-heritage domain thesecould include classes such as “Artist”,“Art_Style” and “Work”. Possible relations are“has_painted” or “has_style”. This meta-knowledge is formalized in ontologies. Thesecond category is that of the specific instancesof the classes and relations in the domain. Theseinstances are stored in a knowledge base. Herewe can find which artists exist, to which specificstyle they belong and which artworks they haveproduced.

In this thesis, we focus on enriching knowledgebases defined by ontologies. We describemethods that automatically identify newinstances of relations and classes. Since in a lotof domains the target information is alreadyavailable on the WWW in human-readable form,the methods use the WWW as the source fromwhich to extract the target information.Information Extraction (IE) is the standard name


for the task of extracting structured informationfrom unstructured documents.

Current IE methods use extraction techniques thatmake use of extensive Natural Language Processing(NLP) or are based on wrappers that exploit explicitstructures in documents. NLP-based methods try tomodel the natural-language sentences in order tounderstand the syntactic structure of the sentence.Using the syntactic structure, the semanticinformation can be extracted. This approach has thedownside that the learned syntactic model islanguage- and structure-dependent. Wrapper-basedmethods try to extract the target information fromstructures such as tables and lists. The performanceof these methods is also dependent on theavailability of such structures in the documents.

In this thesis, we introduce a number of methodsthat are designed to be less dependent on languageand document structure. The methods use simplematching techniques to identify potential targetinformation in documents and use simple term co-occurrence statistics to determine the likelihood ofthe information being correct. By using simplematching and co-occurrence methods, the proposedInformation Extraction methods are able to extractand aggregate information from many documents.This increases the robustness of the methods andallows us to exploit the redundancy of informationon the Web. The proposed methods all extract aworking corpus of documents from the Web, forwhich the co-occurrence statistics are calculated.

In chapters 2, 3 and 4 of this thesis, we present IEmethods that extract instances of relations andclasses defined in ontologies. We evaluate thesemethods by having them extract cultural-heritageinstances that are to be used by the MultimediaN E-culture application. The effectiveness of themethods are further evaluated using similar tasks inother domains.

A METHOD FOR EXTRACTING RELATIONINSTANCES

In Chapter 2, we describe a method that extractsinstances of relations (of the form [Subject, relation,Object]) from the WWW. We use a semi-supervised approach, where we assume that wehave a very small collection of example relationinstances which we aim to automatically expand.The method further assumes that all instances of thesubject and object class of the relation are known:with this method, we do not extract new classinstances, only instances of relations.

An example of this task is the extraction ofinstances of the relation [Art_Style, has_artist,Artist]. Given are lists of both art styles and artists

(both are indeed available in the MultimediaN E-culture knowledge base) and a small seed set ofrelation instances ([Post-Impressionism,has_artist, Vincent_van_Gogh], [Post-Impressionism, has_artist, Gauguin], etcetera).The specific task is to expand this list of relationinstances.

The method described in Chapter 2 follows anumber of steps: First a working corpus ofdocuments is created. For this, the Google searchengine is queried with a search term that matchesone subject instance (for example “Post-Impressionism”). A limited number of resultingdocuments is saved in memory. In the next step,the method identifies instances of the relation’sobject (for example: Artists names) in theworking corpus’ documents. The found instancesare part of candidate relation instances. Next, forevery document, a Document Score isdetermined: the total number of object instancesthat are in the seed set identified in thatdocument divided by the total number of objectinstances in that document. This score representsthe level to which the target relation isrepresented in the document. For each candidateobject instance, the Document Scores are addedand normalized, resulting in an Instance Score.This Instance Score is the likelihood that acandidate relation instance is correct. Therelation instance with the highest associatedInstance Score is added to the seed set, at whichpoint all scores are re-calculated. The iterativeprocess expands the seed set of relation instancesstep-by-step and this makes it possible to startwith a small seed set: in the beginning, the ‘easy’correct expansions will be extracted. As thenumber of iterations grows, so does the seed setand new instances will be identified using moreand more knowledge. To stop this iterativeprocess, we introduce the Drop Factor: when weobserve a relative drop in the Instance Scorebelow some threshold value, the method haltsand presents the current seed set as its finalresult.

We evaluated this so-called ‘redundancymethod’ using experiments in two domains. Inthe first experiment, we extracted instances ofthe relation [Art_Style, has_artist, Artist]introduced earlier. We extracted relationinstances for ten different art styles starting withinitial seed sets of three artists. We manuallyevaluated 400 extracted relation instances. Theresults show that the Drop Factor allows us toselect either higher precision (between 0.62 and0.92) or recall (between 55 and 247 correctrelation instances). We compared the rankingsfor the 400 relation instances based on our


Instance Score with those of a popular semanticdistance measure, the Normalized Google Distance.The results showed that the ranking based onInstance Scores is considerably better. In a secondexperiment, we used the method to extract instancesof the relation [Football_Club, has_player,Football_Player]. The results of this experimentshowed that the method achieves the sameperformance level as for the first experiment. Theoptimal drop factor was also the same for the twotasks.

EXTRACTING TIME PERIODSIn Chapter 3, we describe a method for a morespecific extraction task. Here the goal is to extracttime periods for (historical) concepts. In theMultimediaN E-culture vocabularies, this temporalinformation is not available (for example, for artstyles), even though such information can be ofgreat value.

The method that we present has two main phases. Inthe first phase, we first extract a working corpus ofdocuments by querying the Google search enginewith the target concept’s label. The method thenidentifies potential occurrences of years (four digitnumbers) in all documents. The frequencydistribution of these years is normalized by thefrequency distribution of years on the whole Web.After this normalization step, we are left with afrequency distribution that is representative of thesearch term. We fit a Normal distribution to thisdata. The parameters of this model allows us todetermine the start and end years of a time period.

We evaluated this first phase by extracting timeperiods for art styles, wars, historical periods andartists. We compared the results to a manuallycreated gold standard. The results show a relativelylow error that is also stable for the four types ofconcepts.

In the second phase the extracted time periods,represented by the model parameters aretransformed into instances of a structuredvocabulary. The structured vocabulary used is theTIMEX2 annotation standard that provides anormalized representation of time expressions. InChapter 3, we present a number of rewriting rulesthat convert a frequency model into a TIMEX2instance. We evaluated these rules by applyingthem to the extracted periods for the instances of thefour classes previously introduced. Manualevaluation shows that 88.75% of the TIMEX2periods where considered either completely orpartly correct. Again, this performance was stablefor the different concept types.

PATTERN SPECIFICITY FOR EXTRACTIONPATTERNS

In Chapters 2 and 3, we use very simple, generalmethods that are able to extract information fromdifferent varying documents. In Chapter 4, wetake a closer look at the effect of using moregeneral or more specific extraction methods. Wedo this by introducing a extraction method thatuses a text-analysis application, tOKo. WithtOKo, we can search in corpus documents usingpatterns, in which semantic classes can be used.Using these, we can extract relation instancessuch as [Art_Style, has_artist, Artist] from thedocuments.

To test the effect that the specificity of thepatterns has on the performance of the extractionmethod, we use patterns of varying specificity.Patterns that are more specific make less errorsand therefore result in a higher precision, whilemore general patterns result in a higher recall,usually at the cost of precision. However, whenmore general patterns are used, information fromdifferent documents can be aggregated. When athreshold value on the frequency of pattern hitsis used, a high precision can be attained. Overall,when measuring the performance using the F-measure, using general patterns produces betterresults. The patterns were evaluated in multipledomains, for multiple relation instantiation tasks.

COMBINING BACKGROUND KNOWLEDGE ANDINFORMATION SOURCES

In the evaluation of the previously describedmethods, we found that the use of backgroundknowledge for filtering out faulty candidateinstances can raise the performance of the overallextraction task. A second improvement can beobtained through the combination of multipleinformation sources. This corresponds to the ideain those chapters that using redundancy andmultiple sources is beneficial to the performanceof IE methods.

In Chapter 5 we present a method that usesbackground knowledge and multiple informationsources to lower the likelihood of faultycandidate relation instances (and to effectivelyfilter them out) on one hand and to raise thelikelihood of correct candidate instances on theother hand. We use background knowledge rulesthat produce a likelihood score for a candidaterelation instance using the backgroundknowledge available in the ontology andknowledge base. In Chapter 5, we present anumber of example rules.

One of these rules states for example that it isunlikely that a relation of the type


“participates_in” holds between a subject and anobject that have non-overlapping time periods (e.g.it is unlikely that an artist belongs to an artstyle thatstarted after his death). The rules generate alikelihood score. The different information sourcesalso produce a likelihood score for a candidaterelation instance. In Chapter 5, we discuss a numberof methods through which all these scores can becombined.

Through a number of experiments, we show thatcombining information from different sources withavailable background knowledge indeed raises thequality of a set of candidate relation instances.Errors that are made by one method can becorrected by a different source of information suchas temporal or spatial background knowledge. Moreincorrect candidate relation instances are added tothe knowledge base and the scores of moreincorrect instances drop below the threshold and arefiltered out.

CONCLUSIONOne observation that we made from the evaluationof the methods presented in this thesis is that thesimple, general extraction methods do indeedextract information from different sources. Byaggregating this information from heterogeneoussources and using threshold values on frequenciesof occurrences, good precision scores can still beobtained. We here use the redundancy ofinformation on the Web to our advantage. Thesimple extraction methods are indeed lessdependent on the language and structure of thedocuments and can be re-used for differentextraction tasks in different domains. Evaluationshowed that the performance for the methods isstable in different domains. By using multiplesources, more information can be extracted, leadingto a higher recall. At a meta-level, multipleInformation Extraction sources can also becombined, leading to even more exploitation ofredundancy. In combination with availablebackground knowledge, this can also be used tofilter candidate information, in order to achievehigher recall and precision.

Monte-Carlo Tree Search

Ph.D. thesis abstractGuillaume Chaslot

Promotor: Prof.dr. G. WeissCopromotores: Dr. M.H.M. Winands, dr. B. Bouzy,

dr.ir. J.W.H.M. UiterwijkDate of defense: September 30, 2010

This thesis studies the use of Monte-Carlosimulations for tree-search problems. TheMonte-Carlo technique we investigate is Monte-Carlo Tree Search (MCTS). It is a best-firstsearch method that does not require a positionalevaluation function in contrast to search.MCTS is based on a randomized exploration ofthe search space. Using the results of previousexplorations, MCTS gradually builds a game treein memory, and successively becomes better ataccurately estimating the values of the mostpromising moves. MCTS is a general algorithmand can be applied to many problems. The mostpromising results so far have been obtained inthe game of Go, in which it outperformed allclassic techniques. Therefore Go is used as themain test domain.

Chapter 1 provides a description of the searchproblems that we aim to address and the classicsearch techniques which are used so far to solvethem. The following problem statement guidesour research.

Problem statement: How can we enhanceMonte-Carlo Tree Search in such a way thatprograms improve their performance in a givendomain?

To answer the problem statement we haveformulated five research questions. They dealwith (1) Monte-Carlo simulations, (2) thebalance between exploration and exploitation,(3) parameter optimization, (4) parallelization,and (5) opening-book generation.

Chapter 2 describes the test environment toanswer the problem statement and the fiveresearch questions. It explains the game of Go,


which is used as the test domain in this thesis. Thechapter provides the history of Go, the rules of thegame, a variety of game characteristics, basicconcepts used by humans to understand the game ofGo, and a review of the role of Go in the AIdomain. The Go programs MANGO and MOGO,used for the experiments in the thesis, are brieflydescribed.

Chapter 3 starts with discussing earlier researchabout using Monte-Carlo evaluations as analternative for a positional evaluation function. Thisapproach is hardly used anymore, but it establishedan important step towards MCTS. Subsequently, ageneral framework for MCTS is presented in thechapter. MCTS consists of four main steps: (1) Inthe selection step the tree is traversed from the rootnode until we reach a node, where we select a childthat is not part of the tree yet. (2) Next, in theexpansion step a node is added to the tree. (3)Subsequently, during the simulation step moves areplayed in self-play until the end of the game isreached. (4) Finally, in the backpropagation step,the result of a simulated game is propagatedbackwards, through the previously traversed nodes.

Each step has a strategy associated that implementsa specific policy. Regarding selection, the UCTstrategy is used in many programs as a specificselection strategy because it is simple to implementand effective. A standard selection strategy such asUCT does not take domain knowledge into account,which could improve an MCTS program evenfurther. Next, a simple and efficient strategy toexpand the tree is creating one node per simulation.Subsequently, we point out that building asimulation strategy is probably the most difficultpart of MCTS. For a simulation strategy, twobalances have to be found: (1) between search andknowledge, and (2) between exploration andexploitation. Furthermore, evaluating the quality ofa simulation strategy has to be assessed togetherwith the MCTS program using it. The bestsimulation strategy without MCTS is not always thebest one when using MCTS. The backpropagationstrategy that is the most successful is taking theaverage of the results of all simulated games madethrough a node.

Finally, we give applications of MCTS to differentdomains such as Production Management Problems,Library Performance Tuning, SameGame, MorpionSolitaire, Sailing Domain, Amazons, Lines ofAction, Chinese Checkers, Settlers of Catan,General Game Playing, and in particular Go.

The most basic Monte-Carlo simulations consist ofplaying random moves. Knowledge transforms theplain random simulations into more sophisticated

pseudo-random simulations. This has led us tothe first research question.

Research question 1: How can we useknowledge to improve the Monte-Carlosimulations in MCTS?

Chapter 4 answers the first research question.We explain two different simulation strategiesthat apply knowledge: urgency-based andsequence-like simulation. Based on theexperience gathered from implementing them inINDIGO and MOGO, respectively, we make thefollowing three recommendations. (1) Avoidingbig mistakes is more important than playinggood moves. (2) Simulation strategies usingsequence-like simulations or patterns in urgency-based simulations are efficient because theysimplify the situation. (3) The simulationstrategy should not become too stochastic, nortoo deterministic, thus balancing exploration andexploitation.

Moreover, we develop the first efficient methodfor learning automatically the knowledge of thesimulation strategy. We proposed to use moveevaluations as a fitness function instead oflearning from the results of simulated games. Acoefficient is introduced that enables to balancethe amount of exploration and exploitation. Thealgorithm is adapted from the tracking algorithmof Sutton and Barto. Learning is performed for9×9 Go, where we showed that the Go programINDIGO with the learnt patterns performed betterthan the program with expert patterns.

In MCTS, the selection strategy controls thebalance between exploration and exploitation.The selection strategy should favour the mostpromising moves (exploitation). However, lesspromising moves should still be investigatedsufficiently (exploration), because their lowscores might be due to unlucky simulations. Thismove-selection task can be facilitated byapplying knowledge. This idea has guided us tothe second research question.

Research question 2: How can we useknowledge to arrive at a proper balance betweenexploration and exploitation in the selection stepof MCTS?

Chapter 5 answers the second research questionby proposing two methods that integrateknowledge into the selection step of MCTS:progressive bias and progressive widening.Progressive bias uses knowledge to direct thesearch. Progressive widening first reduces thebranching factor, and then increases it gradually.


We refer to them as “progressive strategies”because the knowledge is dominant when thenumber of simulations is small in a node, but losesinfluence progressively when the number ofsimulations increases.

First, the progressive strategies are tested inMANGO. The incorporated knowledge is based onurgency-based simulation. From the experimentswith MANGO, we observe the following. (1)Progressive strategies, which focus initially on asmall number of moves, are better in handling largebranching factors. They increase the level of play ofthe program MANGO significantly, for every boardsize. (2) On the 19×19 board, the combination ofboth strategies is much stronger than each strategyapplied separately. The fact that progressive biasand progressive widening work better incombination with each other shows that they havecomplementary roles in MCTS. This is especiallythe case when the board size and thereforebranching factor grows. (3) Progressive strategiescan use relatively expensive domain knowledgewith hardly any speed reduction.

Next, the performance of the progressive strategiesin other game programs and domains is presented.Progressive bias increases the playing strength ofMOGO and of the Lines-of-Action program MC-LOA, while progressive widening did the same forthe Go program CRAZY STONE. In the case ofMOGO, progressive bias is successfully combinedwith RAVE, a similar technique for improving thebalance between exploitation and exploration.These results give rise to the main conclusion thatthe proposed progressive strategies are essentialenhancements for an MCTS program.

MCTS is controlled by several parameters, whichdefine the behaviour of the search. Especially theselection and simulation strategies contain severalimportant parameters. These parameters have to beoptimized in order to get the best performance outof an MCTS program. This challenge has led us tothe third research question.

Research question 3: How can we optimize theparameters of an MCTS program?

Chapter 6 answers the third research question byproposing to optimize the search parameters ofMCTS by using an evolutionary strategy: the Cross-Entropy Method (CEM). CEM is related toEstimation-of-Distribution Algorithms (EDAs), anew area of evolutionary computation. The fitnessfunction for CEM measures the winning rate for abatch of games. The performance of CEM with afixed and variable batch size is tested by tuning 11parameters in MANGO. Experiments reveal that

using a batch size of 500 games gives the bestresults, although the convergence is slow. Asmall (and fast) batch size of 10 still gives areasonable result when compared to the best one.A variable batch size performs a little bit worsethan a fixed batch size of 50 or 500. However,the variable batch size converges faster than afixed batch size of 50 or 500.

Subsequently, we show that MANGO with theCEM parameters performs better against GNUGO than the MANGO version without. In fourself-play experiments with different time settingsand board sizes, the CEM version of MANGOdefeats the default version convincingly eachtime. Based on these results, we may concludethat a hand-tuned MCTS using game engine mayimprove its playing strength when re-tuning theparameters with CEM.

The recent evolution of hardware has gone intothe direction that nowadays personal computerscontain several cores. To get the most out of theavailable hardware one has to parallelize MCTSas well. This has led us to the fourth researchquestion.

Research question 4: How can we parallelizeMCTS?

Chapter 7 answers the fourth research questionby investigating three methods for parallelizingMCTS: leaf parallelization, root parallelizationand tree parallelization. Leaf parallelizationplays for each available thread a simulated gamestarting from the leaf node. Root parallelizationconsists of building multiple MCTS trees inparallel, with one thread per tree. Treeparallelization uses one shared tree from whichgames simultaneously are played.

Experiments are performed to assess theperformance of the parallelization methods in theGo program MANGO on the 13×13 board. Inorder to evaluate the experiments, we proposethe strength-speedup measure, whichcorresponds to the time needed to achieve thesame strength. Experimental results indicate thatleaf parallelization is the weakest parallelizationmethod. The method leads to a strength speedupof 2.4 for 16 processor threads. The simple rootparallelization turns out to be the best way forparallelizing MCTS. The method leads to astrength speedup of 14.9 for 16 processorthreads. Tree parallelization requires twotechniques to be effective. First, using localmutexes instead of a global mutex doubles thenumber of games played per second. Second, thevirtual-loss enhancement increases both the


games-per-second and the strength of the programsignificantly. By using these two techniques, weobtain a strength speedup of 8.5 for 16 processorthreads.

Modern game-playing programs use opening booksin the beginning of the game to save time and toplay stronger. Generating opening books incombination with an program has been wellstudied in the past. The challenge of generatingautomatically an opening book for MCTS programshas led to the fifth research question.

Research question 5: How can we automaticallygenerate opening books by using MCTS?

Chapter 8 answers the fifth research question bycombining two levels of MCTS. The method iscalled Meta Monte-Carlo Tree Search (Meta-MCTS). Instead of using a relatively simplesimulation strategy, it uses an entire MCTSprogram (MOGO) to play a simulated game. Wedescribe two algorithms for Meta-MCTS: QuasiBest-First (QBF) and Beta-Distribution Sampling(BDS). The first algorithm, QBF, is an adaptation ofgreedy algorithms that are used for the regularMCTS. QBF favours therefore exploitation. Duringactual game play we observe that despite the goodperformance of the opening book, some branchesare not explored sufficiently. The second algorithm,BDS, favours exploration. In contrast to UCT, BDSdoes not need an exploration coefficient to be tuned.The algorithm draws a move according to itslikelihood of being the best move (considering thenumber of wins and losses). This approach createdan opening book which is shallower and wider. TheBDS book has the drawback to be less deep againstcomputers, but the advantage is that it stayed longerin the book in official games against humans.Experiments on the Go server CGOS reveal thatboth QBF and BDS were able to improve MOGO. Inboth cases the improvement is more or less similar.Based on the results, we may conclude that QBFand BDS are able to generate an opening bookwhich improves the performance of an MCTSprogram.

The last chapter of the thesis returns to the fiveresearch questions and the problem statement asformulated in Chapter 1. Taking the answers to theresearch questions above into account we see thatthere are five successful ways to improve MCTS.First, learning from move evaluations improves theknowledge of the simulation strategy. Second,progressive strategies enhance the selection strategyby incorporating knowledge. Third, CEM optimizesthe search parameters of an MCTS program in sucha way that its playing strength is increased. Fourth,MCTS benefits substantially from parallelization.

Fifth, Meta-MCTS generates an opening bookthat improves the performance of an MCTSprogram. Yet, we are able to provide additionalpromising directions for future research. Finally,the question of understanding the nature ofMCTS is still open.

Converting and IntegratingVocabularies for the Semantic Web

Ph.D. thesis abstractMark van Assem

Promotor: Prof.dr. A.Th. SchreiberCopromotor: Dr. J.R. van OssenbruggenDate of defense: October 1, 2010

Institutions such as libraries, museums and otherarchives have been collecting books, paintings,statues and other objects for centuries. Tomanage these collections, cataloguers havedescribed each object with respect to its title,author, subjects, materials and other attributes.This process is called “indexing”, and simplifiesthe process of searching through the collections.An object description created during indexing isessentially a set of attribute-value pairs. Such adescription might consist e.g. of pairsauthor=Rembrandt, title=Anatomy Lesson,date=1632, type=painting, subject=groupportrait. Such descriptions are also calledmetadata (data about the actual object).

The values for the pairs are usually taken fromvocabularies. Vocabularies are lists of conceptswith definitions and play a key role in indexing


and search. Firstly, they offer a set of agreed uponconcepts that cataloguers can pick from. Secondly,concepts provide a convenient place to groupsynonymous terms (e.g. “clair-obscure” and“chiaroscuro” which both refer to Rembrandt’spainting style). Thirdly, the concepts usually have aunique identifier, which allows the cataloguer toindicate the correct concept even though theconcept has an ambiguous term (e.g. “painting” asthe process of applying a protective coating to anobject vs. “painting” as the process of creating anexpressive or communicative image). Fourthly, theconcepts are often placed into a hierarchy (e.g.“origami” below “Japanese art”) which simplifiessearch (a search for books on Japanese art will alsoreturn books on origami).

To aid the indexing process, each institution notonly prescribes a set of vocabularies to be used, butalso the attributes and their names. Attributes arecalled elements, and the set of allowed attributes istogether called the metadata element set.

With the advent of the Web people and institutionshave started sharing their data. This has thepotential benefit that all information on a particulartopic, say the paintings of Vincent van Gogh, can bequeried as if they were stored in one system. Thisrequires that the data is integrated: it must conformto a particular format and structure that the searchsystem understands. Two obstacles to integrationare the different data formats in use (the syntacticintegration problem) and the different terms in useto denote similar concepts (the semantic integrationproblem). One example of the latter problem iswhen one institution has a concept called “clair-obscure” and another has a concept called“chiaroscuro”. Another example is when oneinstitution uses a metadata element called“creator” while another has an element called“author”. These concepts and elements are highlysimilar and a search for clair-obscure paintings byRembrandt needs to query through both conceptsand elements.

The Semantic Web is a research area that proposesparticular solutions to these problems. Firstly, itproposes to use a family of web-based knowledgerepresentation languages that have RDF asunderlying model (RDFS and the several flavoursof OWL). Conversion of data to this family oflanguages solves a substantial part of the syntacticintegration problem. Secondly, the languages have afew simple mechanisms to relate similar concepts toeach other, solving part of the semantic integrationproblem. For example, it is possible to state that“creator” is equivalent to “author”, so that a queryon either element will automatically include theresults obtained from querying with the other.

In this thesis we assume that the approach andlanguages proposed by the Semantic Webcommunity are useful for achieving integration,and aim to apply these in the context of thecultural-heritage domain. The problem ofconverting the original data sets to RDF/OWLhas not been investigated much. In this thesis wefocus mostly on conversion of vocabularies. Ourproblem statement is as follows: How canexisting vocabularies be made available toSemantic Web applications? Problems that needto be solved include understanding the originalsyntactical format in which the vocabulary isexpressed, understanding the conceptual modelthat lies behind it, linking this conceptual modelto that of RDF/OWL, and finding an appropriateway to convert the former model into the latter.These tasks are far from being automated, butthe demand for proper conversions will increasein the coming years. Therefore, this thesis hasfocused on developing methods for conversion ofvocabularies. Methods are step-wise processeswith guidelines that can be followed by peopleperforming the conversion task. Several choiceshave to be made during the process that affectthe resulting representation. Conversion can beperformed for the benefit of one particularapplication and tuned to its specific needs, butconversions can also aim at a representation thatis as complete and reusable as possible for anyapplication. The main contribution of this thesisis the development of two separate methods tocater to both situations.

In Chapter 2 we develop a first version of ageneric method for conversion. The assumptionis that a faithful and complete conversion of thevocabularies results in a representation that isuseful for most applications. A methodconsisting of several steps and guidelines wasdrafted, and then applied to two case studies:conversions of the MeSH and WordNetvocabularies. These helped to improve themethod; they showed which additionalguidelines were needed to adequately handlethese cases. We deliberately chose two complexvocabularies so that a broad range of vocabularyfeatures were covered. Two tailor-made schemasfor each vocabulary and a complete conversionof their content to these schemas are theoutcome.

Another way to convert vocabularies suitable fora wide range of applications is to use a standard,widely supported vocabulary schema. In Chapter3 we developed a method aimed at the emergingSKOS standard. We chose three vocabularies asuse cases: a simple one (GTAA), an


intermediately complex one (IPSV) and a complexcase (MeSH). We found that SKOS was suitable forconverting GTAA and IPSV (making use ofRDF/OWL abilities to specialize a schema), butMeSH could not be covered completely becauseSKOS did not allow a concept’s terms to berepresented as instances themselves.

In Chapter 4 we returned to the problem of genericconversion as approached in Chapter 2. Theassumption that our method results in vocabulariesuseful for many applications was tested bycomparing our generic conversion of WordNet to aconversion developed with application use cases inmind (in the W3C Semantic Web Best PracticesWorking Group). The comparison of the twoWordNets showed how our generic method shouldbe changed to cater for a wider range ofapplications. However, another outcome of thechapter is that a generic method cannot cater to allrequirements an application might have, because itmay require that content is left out or structureddifferently than in the original source. We alsoimproved the WordNet conversion to SKOS byapplying a newly developed extension that allowsterms to be represented as instances themselves.This solves most of the problems noted in theMeSH conversion to SKOS.

Given the results from Chapter 4, we developed anew method that can be used to cater to specificapplications in Chapter 5. We adapted the genericmethod by introducing specific steps to determinethe requirements and use cases that need to becovered. The principle of complete and faithfulconversion of the original source was dropped. Thecase study is the MultimediaN E-Culture search andbrowsing application, for which the AAT, TGN andULAN were converted. This case study pointed outthat the conversion made by the E-Culture teamwithout our method missed several pieces ofinformation needed by the application use cases.

In Chapter 6 we continued to investigateconversions targeted at specific applications. In thischapter we concentrated on alignment applications,which take two or more vocabularies in RDF/OWLas input and produce mapping relations betweenconcepts of the vocabularies. Our analysis showedthat these applications cannot handlerepresentations as produced by our methods. Weprovided a conversion technique to mitigate thisproblem. The analysis is part of a study on newevaluation techniques for alignments ofvocabularies. Alignment is a central ingredient ofintegration as promoted by the Semantic Webcommunity. The outcome of the study is that ourproposed techniques are better tuned to evaluating

the quality of an alignment for a particularapplication than existing techniques.

Integration of collections relies on integration ofboth vocabularies and metadata element sets. InChapter 7 we study how an existing metadataelement set can be represented in RDF/OWL in away that is interoperable with vocabularyrepresentations as advocated in this thesis. Themetadata element set is called VRA and catersspecifically to cultural heritage. We show howVRA can be implemented as a specialization ofthe more generic Dublin Core element set.Linking VRA with Dublin Core allowsintegration of collections from different domains(television archives, libraries, etcetera) into onesearch system. We also show how VRA can bespecialized to reflect that e.g. the Rijksmuseumuses e.g. ULAN as range of the “creator”element (we term this feature collection-specificvalue ranges).

In summary, this thesis contributes to theintegration of metadata collections in three ways.Firstly and chiefly through the development ofconversion methods for vocabularies andthrough contributing actual conversions madewith the methods. Secondly, throughinvestigating how metadata schemas can berepresented in a way that allows using themtogether with vocabulary representationsproduced by the methods. Thirdly, bycontributing a study on how alignments can beevaluated on their usefulness for particularapplications.

A Computational Approach toContent-Based Retrieval of Folk Song

Melodies

Ph.D. thesis abstractPeter van Kranenburg

Promotores: Prof.dr. R.C. Veltkamp, Prof.dr.L.P. Grijp

Copromotor: Dr. F. WieringDate of defense: October 4, 2010


In order to develop a Music Information Retrievalsystem for folk song melodies, one needs to designan adequate computational model of melodicsimilarity, which is the subject of this Ph.D. thesis.Since fundamental understanding of melodies inoral culture as well as fundamental understanding ofcomputational methods to model similarity relationsbetween folk song melodies both are necessary, thisproblem needs a multidisciplinary approach.

Chapter 2 reviews the relevant academicbackground of both Folk Song Research (as sub-discipline of Ethnomusicology) and MusicInformation Retrieval. It also presents aninterdisciplinary collaboration model in whichComputational Musicology serves a ‘man-in-the-middle’ role. The particular task of ComputationalMusicology is to design computational models ofconcepts from Musicology. In the context of thisthesis, the concept of tune family is the mostimportant.

An important step towards the understanding of theconcept of tune family is the method to annotatesimilarity relations between melodies that ispresented in Chapter 3. It has been developed inclose collaboration with musicological domainexperts to make aspects of their intuitive similarityassessments explicit. 360 melodies in 26 tunefamilies were ‘manually’ annotated, resulting in anAnnotated Corpus that is a valuable resource for thestudy of melodic similarity and for the evaluation ofcomputational models of melodic similarity. Fromthe annotations we conclude that the relativeimportance of the various dimensions (global andlocal rhythm, global and local contour, motifs,lyrics) varies to a large extent in individualcomparisons. Furthermore, the dimension of motifs

seems the most important for recognizingmelodies. This means that in many casesmelodies are judged to be related based onshared characteristic melodic motifs.

In Chapter 4, 88 low-level, global, quantitativefeatures of melody are used to discriminatebetween tune families. It appears that suchfeatures can be used to recognize melodieswithin a relatively small corpus (such as the 360melodies in the Annotated Corpus), but thatthese features lose their discriminative power ina larger dataset of thousands of melodies.

Chapter 5 uses the same kind of features inanother musical domain: baroque organ fugues.Global features are used to assess authorshipproblems of fugues that are in the catalog of J.S.Bach. Several hypotheses from musicologicalliterature about the authorship of several fuguesare supported by findings using this method. Thevarious degrees of success of the samecomputational method in the previous and thecurrent chapters show that computationalmethods cannot ‘blindly’ be applied tomusicological questions.

In Chapter 6, the potential of alignmentalgorithms for folk song melody retrieval isstudied by incorporating musical knowledge inthe algorithm in the form of appropriate,musically motivated, substitution scoringfunctions. This approach leads to good retrievalresults both for a small (360 melodies) and alarge (4830 melodies) dataset. Furthermore,domain experts were able to classify‘problematic’ melodies using the results of thealignment algorithms.

The final chapter introduces a local, motif-basedapproach to the retrieval of folk song melodiesusing audio recordings rather than the symbolictranscriptions. The melodies are segmented bydetecting breath and pauses. Using the resultingaudio segments, it proves possible for most ofthe tune families in the Annotated Corpus to finda characteristic segment that has approximatelocal matches in other melodies from the sametune family. Retrieval results using thesemusically meaningful segments are better thanusing fixed-length segments. The overallretrieval results are not good enough to use in anend user retrieval system yet. However, there ismuch room for improvement.

This thesis contributes both to Folk SongResearch and Music Information Retrieval byincorporating musical knowledge incomputational models. The process of


developing such models of similarity relationsbetween folk song melodies leads to betterunderstanding of melodic similarity and, thus, of theconcept of tune family, which is relevant for FolkSong Research. The models that have beendeveloped, have successfully been used in theretrieval experiments. From the research that ispresented in this thesis, it is clear that computationalmethods have a rich potential for the study ofmusic; not as a replacement of ‘traditional’methods, but as an extension of the researchmethods that are available to the musicologist.

Allograph Based Writer Identification,Handwriting Analysis and Character

Recognition

Ph.D. thesis abstractRalph Niels

Promotores: Prof.dr. L.R.B. Schomaker, Prof.dr.ir.P.W.M. Desain

Copromotor: Dr. L.G. VuurpijlDate of defense: October 4, 2010

In this thesis, techniques and features are describedthat were developed for the automatic comparisonof handwritten characters by first matching them toprototypical character shapes (allographs). Thesetechniques and features were evaluated inexperiments simulating different real-worldapplications.

The majority of the experiments regard forensicwriter identification, where the objective is to findthe writer of a piece of handwriting by comparing it

to a large set of handwritten documents of whichthe writer is already known. The assumption isthat if two documents contain many similarallographs, they may have been produced by thesame writer. In the experiment described, it isdemonstrated that using the techniques andfeatures, it is indeed possible to match thecorrect writer with a piece of unknownhandwriting.

Other experiments were performed to evaluatethe usefulness of the techniques and features forthe classification of hand-drawn symbols andcharacters in different scripts (where theobjective is not to find out who produced thewriting, but what it represents) and the analysisof children’s handwriting to diagnoseDevelopmental Coordination Disorder.

Ecological Automation Design,Extending Work Domain Analysis

Ph.D. thesis abstractMatthijs Amelink

Promotores: Prof.dr.ir. M. Mulder, Prof.dr.ir. B.Mulder

Copromotor: Dr.ir. R. van PaassenDate of defense: October 18, 2010

In high-risk domains like aviation, medicine andnuclear power plant control, automation hasenabled new capabilities, increased the economyof operation and has greatly contributed tosafety. However, automation increases thenumber of couplings in a system, which caninadvertently lead to more complexity from the


perspective of the operator. The automation of asystem transforms the work domain of the humanoperator, and his role changes from controlling thecore processes to managing the automatedprocesses. The complexity of the automation andthe lack of proper support can make the controltask’s overall difficulty larger than it needs to be,restricting safety, productivity, and efficiency.

To address and limit the automation introducedcomplexity in the operator’s work domain, and tofind representations to support him, the ecologicalapproach to automation design was taken. Theecological approach focuses on the relationshipbetween the human operator and his work domainincluding the system he is controlling. The mainresearch goals were to find how the ecologicalapproach could be used to help limit the automationintroduced complexity, and how the ecologicalapproach could be used to support the humanoperator in controlling automated processes.

The formulation of Ecological Automation Design(EAD) was based on the Ecological InterfaceDesign (EID) paradigm. One of the mainunderlying questions asked about the interfacebetween the work domain and the human operatoris: “how to represent work domain complexity?".The interface design paradigm was transformed intoan automation design paradigm by first separatingthe automation component from the work domainand asking the same underlying question about theinterfaces between the work domain, the humanoperator, and the automation. Then, the conceptualshared domain representation was defined tovisualize that the apparent complexity of the systemcould be reduced when both the human operatorand the automation view the same representation ofconstraints that the work domain imposes oncontrol. As part of the ecological approach, WorkDomain Analysis (WDA) was used to analyze andrepresent the constraints in a work domain.However, WDA is not yet fully developed andsuffers from some methodological and conceptualissues. The research therefore focused on the furtherdevelopment and extension of WDA to include therepresentation of automated processes. Four casestudies were conducted, and each case studygenerated new insights into the application of andextension of WDA.

In the first case study, EID was applied to thedesign of the Energy Augmented Tunnel In the Skydisplay. This display was designed to aid a pilot tofly the approach to landing by presenting energymanagement information. The WDA revealed thesignificance of the energy coupling between verticalflight path and speed control as an intermediatecontrol goal. Based on the analysis, a creative

design process resulted in a novel display thathas the energy representations fully, andgraphically integrated in the tunnel in the skydisplay. A preliminary evaluation indicated thatthe additional energy management informationshown in relation to the control actions andcontrol goals helped pilots to fly the approaches.The display is not expected to give aperformance increase but to change the way inwhich pilots control the throttle and elevator tofly approaches.

The second case study was the analysis of thealready existing Total Energy Control System(TECS). TECS is an unconventional automatedflight control system that was based on the sameenergy management constraints as that wererepresented in the energy augmented display ofthe first case study. The design of TECS wasmapped onto the abstraction hierarchy torepresent the energy management principles aspart of the whole automated system. The analysisand useful representation of TECS using theabstraction hierarchy was not straightforward. Itinvolved a search for the interpretation of thelevels of the abstraction hierarchy and the use ofthe means-ends relationship in conjunction withthe aggregation relationship. The resulting WDAshowed that the abstraction hierarchy could beused to map out the reasons for TECS’s designfeatures. Many constraints were represented inthe same space, which cluttered the energymanagement principles. The focus was put onthe energy management principles throughselective aggregation of the representedfunctions, but other design principles wereomitted. To provide a complete representation ofthe system but without the clutter, the levels ofcontrol sophistication were introduced torepresented nested control problems separately.At each level of control sophistication theabstraction hierarchy was applied, resulting inthe Abstraction-Sophistication Analysis (ASA).

In the third case study, the ASA framework wasused to guide the design of SmartUAV.SmartUAV is a newly designed mini–UAVsystem that is capable of controlling multiplesmall UAVs from a laptop computer. Bydesigning and developing SmartUAV we gainedhands-on experience with how WDA, andespecially the ASA, helped to keep track of anddeal with the automation introduced constraintsin the design phase. The levels of controlsophistication were used from the beginning toseparate the different control problems in thedomain. They ranged from flying the platform tothe achievement of missions. Starting at thelowest level of control sophistication, each


higher level allowed the designer to include a largerpart of the complete work domain incrementally,and to focus on more sophisticated control of theUAV. Furthermore, the ASA supported thevisualization of how automation transformed thework domain, thus how automated functionalitiesthat were created at lower levels of controlsophistication affected the (automated) functions athigher levels of control sophistication. This studyshowed that the ASA could span a much largerproblem space than the original WDA through thenesting of abstraction hierarchies. The ASAprovided a systematic way to address theabstraction of the control problems (levels ofcontrol sophistication) and the abstraction offunctions per control problem (abstractionhierarchy).

The fourth case study dealt with the analysis of asubset of a well-structured domain that lacksautomation; sailboat racing. This study generated aclearer view on the nested structure that is inherentin a work domain, as opposed to the nestedstructure of the automation as found in TECS andSmartUAV. The nested structure inherent to thiswork domain was found to be the result of howsailboat racing has evolved over time, based on thecapabilities of equipment, human performance andthe racing rules. Due to the lack of automation, itbecame clear that human performance is in fact partof the work domain, in contrast to the originalformulations of WDA. The crew’s performanceformed the basis for achieving the moresophisticated control of boat speed, tactics andstrategy, thus was essential in the analysis. It wasshown that the performance of the human crewcould be represented in the ASA at a level ofcontrol sophistication, while this could not besupported in a non-nested WDA based on a singleabstraction hierarchy.

The four case studies exemplified WDA and led toits extension with a structure to explicitly nestabstraction hierarchies that map out differentcontrol problems: the ASA. Through generating theanalyses, extensive modeling experience with theabstraction hierarchy was obtained, reducing itsambiguity and potential methodological andconceptual problems. We found that the abstractionhierarchy could be used to model the structure ofthe knowledge about a work domain but could notmodel the knowledge itself. Therefore, theabstraction hierarchy is a framework for structuringknowledge, linking different representations of acontrol problem, and explaining the reasons fordesign features of a system.

The abstraction hierarchy addressed the abstractionof elements belonging to a control problem, and the

levels of control sophistication addressed theabstraction of the control problem itself.Representations in the ASA framework rangedfrom physical at the lower levels of controlsophistication to non-physical at the higherlevels of control sophistication. It allowed thestructuring of, for example, the sailboat racingrules at the higher levels, and the law ofconservation of energy at the lower levels.Although the application of the ASA did notinherently reduce the complexity of the design ofSmartUAV, it enabled us to better understand theelements of the work domain that contribute tocomplexity of the system prior to and during itsdesign. The extension of work domain analysiswith the levels of control sophistication has ledto a richer representation of the studied workdomains than a single abstraction hierarchy orthe abstraction-decomposition space.

Context-Based Sound EventRecognition

Ph.D. thesis abstractMaria Niessen

Promotor: Prof.dr. L.R.B. SchomakerCopromotor: Dr. T.C. AndringaDate of defense: October 22, 2010

The research domain of automatic sound eventrecognition aims to describe an audio signal interms of the sound events that compose a sonicenvironment. The ability to recognize events in areal-world environment requires a listener orsystem to separate the sound events from each


other and the background. Furthermore, theseseparated events need to be recognized. Torecognize a sound event implies that somerepresentation of the event is already known to thereceiver, and can be identified when it isencountered again. The ability of sound eventrecognition depends not only on the audio patternspecific to the event, but also on the semantics ofthe event. For example, a sound of a purring catmay seem unique, but without any otherinformation than represented by the audio signal, itcan sound like an engine as well. In this thesis weshow that the task of recognizing a sound event canbe alleviated with the semantics of the event, whichis inferred from a model of the context in which theevent occurs.

A possible strategy to provide an automatic soundrecognition system with the semantics of a soundevent, is to develop it for a specific application, andhence a specific type of context. For example,automatic speech recognition systems expect aspeech signal as input, which is ensured by a user.Therefore, particular assumptions about the audiosignal can be made, and context information in theform of grammar rules can be applied to recognize aword sequence (section 2.2). However, if a systemfor automatic sound event recognition is notdesigned for a specific application, and should workin variable and uncontrolled real-worldenvironments, no assumptions about theenvironmental conditions can be made. Therefore,additional analysis of the audio signal is required.First, the sound events to be recognized have to beseparated from the background, because the inputsignal consists of more than one type of sound(section 2.3). Second, the operating environmentcannot be controlled, hence the system has to dealwith transmission effects, such as reverberation(section 2.4).

In addition to handling the challenges of real-worldenvironments with signal-driven methods, thesemantics of the event and its environment areessential to recognize sound events in an unreliableor ambiguous audio signal. People have nodifficulty in recognizing sound events in manydifferent and noisy situations. Hence, we developeda model for automatic sound event recognition thatis inspired by the strategies of human listeners asinvestigated by (psycho-)acoustics and cognitivepsychology. The human percept of a sound event,referred to as an auditory event (section 3.2), hasproperties that a representation of a sound event inan automatic system can benefit from as well.People can generalize auditory events over differentexperiences, environments, and senses. Therefore,our model should store invariant representations ofsound events, so that it is robust to variable

environmental conditions, similar to humanperception. Moreover, people benefit frominformation about the environmental context torecognize sound events (section 3.3). Thisfacilitatory effect of contextual knowledge is animportant design objective of our model.

Some other studies have been aimed at modelingcontext awareness in acoustics (section 4.1).However, in these studies the goal has been toestimate the context in itself, rather than to usecontext for the improvement of sound eventrecognition. In other research domains, such asinformation retrieval and handwritingrecognition, context has been used to improverecognition or retrieval of objects. Oftenmethods in these research domains use spreadingactivation networks, which are based on a modelof human memory (Collins and Loftus, 1975). Inthis thesis we show that spreading activationnetworks can also be applied to estimate themost likely interpretation of an audio pattern. Weintroduce a context model in which thesemantics of the events and the context arerepresented as nodes in the network (section4.2). The activation of the nodes spreads throughthe network to determine the confidence ofpossible interpretations of an audio pattern(section 4.3).

The advantage of modeling context in automaticsound event recognition is demonstrated byapplying an integrated system to audio recordedin real-world environments. This integratedsystem is a combination of a signal-drivenanalysis of the audio signal, which provideshypotheses of sound events, and the contextmodel, which interprets these hypotheses.Knowledge about the environmental context islearned in a training phase. This knowledge isrepresented as a network of nodes, in which thenodes represent the semantics of the soundevents and the different contexts. Furthermore,the connections between the nodes carry weightsthat indicate the probabilities that these soundevents and contexts are encounteredsubsequently or concurrently. The values of theweights between nodes are learned fromannotated training data (section 5.2). The type ofcontext that is learned depends on the applicationdomain and the data set. In a stable environment,information of co-occurring events can help toform expectancies of future events. For example,at a train station the beeping of a closing door islikely to be followed by a departing train.Alternatively, if the data set is recorded atqualitatively different types of locations, theestimated location can help to predict the typesof sound events that may be heard. For example,


birds are more commonly heard in parks than nearbusy roads. We explored the benefit of both typesof context on sound event recognition in twoexperiments (sections 5.3 and 5.4). The results ofthese experiments show that the evaluation ofcontextual knowledge improves the recognition ofsound events compared to an exclusively signal-driven method.

Contextual knowledge is not restricted toknowledge that can be derived from the audiosignal and annotations of the audio signal. Othertypes of knowledge can be beneficial for soundevent recognition as well. Knowledge inferred fromdifferent types of input, such as sound, image, andlocation, can reinforce each other to obtain anincreased awareness of events in the environment.Ideally, these different informational resources canbe combined in a single system. Because the nodesin the context model are not described by modalityspecific knowledge, they can be used for other typesof information. We tested the applicability of thecontext model on the recognition of ambiguousvisual information, in the domain of robotlocalization. Visual information received by a robotis often ambiguous (similar to acousticinformation), because similar observations, such as(parts of) chairs or windows, can be made at distinctplaces in an environment. Learned knowledge aboutthe environment, and the robot’s hypothesizedposition in the environment, can help todisambiguate these observations. As a result, theposition prediction improves compared to a signal-driven approach (chapter 6).

The Value of a Thesis

H. Jaap van den HerikTiCC, UvT, Tilburg

The economic decrease encourages researchers toput emphasis on collecting more knowledge, so thatthey have many things to ponder and to deepen intimes of scarcity. The value of a thesis can be seenas an investment in the future. However, it is notclear whether all ideas can be implemented andrealised in the real world. So, next to the value of athesis, we have the value of an application. InScience, we see many times a discussion in theprocedure of approving a thesis. A first prerequisiteis, of course, whether the thesis contains new ideas.Yet, ideas that cannot be followed by a process ofrealisation lose some of their value.

Currently, at many universities we see that the“Breimer” norm has been accepted. This means that

a thesis should contain roughly 100,000 words.The norm is introduced to protect thepromovendus, the supervisor, and the assessmentcommittee against an overload of effort whenreading (for the promovendus when writing) thethesis.

At this moment, the constructive sciences are atopic of study by NWO committees and KNAWcommittees. The prevailing question here iswhether a “construction” (airplane, boat,intelligent bridge, autonomous train) can be seenas a new idea, a contribution to science and apiece of work that is worth the Doctor’s title. Ifso, then the following question is whether theconstruction of a piece of art is also acontribution to cultural sciences and thus worththe Doctor’s title.

In the list of announcements we see the thesis byB. Thomee titled A Picture is Worth a ThousandWords. A simple calculation based on the aboveestimations brings us the following. The“Breimer” norm may lead to dissertationscontaining some one hundred pictures. In theextreme case, we may thus expect a thesis ofhundred pictures. If they have the quality ofRembrandt van Rijn, Jacob Maris, or JacobZekveld it may lead to a Doctor’s title. We willsee what the future has in store.

The current list of announcements is long anddivers. Many titles are long (too long, in myopinion), but I also see titles with five or fewerwords (well done). For all promovendi, amilestone is announced, their thesis. Such anevent deserves our congratulations. We wish alldefenders a bright future.

Mark van Assem (VU) (October 1, 2010).Converting and Integrating Vocabularies for theSemantic Web. Vrije Universiteit Amsterdam.Promotores: Prof.dr. A.Th. Schreiber (VU).Copromotor: Dr. J.R. van Ossenbruggen(VU/CWI).

Vasilios Andrikopoulos (UvT) (October 1,2010). A Theory and Model for the Evolution ofSoftware Services. Tilburg University. Promotor:Prof.dr. M. Papazoglou (UvT).

Peter van Kranenburg (UU) (October 4, 2010).A Computational Approach to Content-BasedRetrieval of Folk Song Melodies. UtrechtUniversity. Promotores: Prof.dr. R.C. Veltkamp(UU), Prof.dr. L.P. Grijp (UU). Copromotor: Dr.F. Wiering (UU).


Ralph Niels (RUG) (October 4, 2010). AllographBased Writer Identification, Handwriting Analysisand Character Recognition. University ofGroningen. Promotor: Prof.dr. L.R.B. Schomaker(RUG). Copromotor: Dr. L.G. Vuurpijl (RUN).

Mark van Staalduinen (TUD) (October 7, 2010).Content-Based Paper Retrieval. Towardsreconstruction of art history. Delft University ofTechnology. Promotor: Prof.dr.ir. J. Biemond(TUD). Copromotor: Dr.ir. J.C.A. van den Lubbe(TUD).

Pieter Bellekens (TU/e) (October 7, 2010). AnApproach towards Context-Sensitive and User-Adapted Access to Heterogeneous Data Sources,Illustrated in the Television Domain. EindhovenUniversity of Technology. Promotores: Prof.dr.P.M.E. De Bra (TU/e), Prof.dr.ir. G.J.P.M Houben(TUD). Copromotor: Dr. L.M. Aroyo (VU).

Matthijs Amelink (TUD) (October 18, 2010).Ecological Interface Design, Extending WorkDomain Analysis. Delft University of Technology.Promotores: Prof.dr.ir. M. Mulder (TUD),Prof.dr.ir. B. Mulder (TUD), Copromotor: Dr.ir. R.van Paassen (TUD).

Maria Niessen (RUG) (October 22, 2010).Context-Based Sound Event Recognition. Universityof Groningen. Promotor: Prof.dr. L.R.B. Schomaker(RUG). Copromotor: Dr. T.C. Andringa (RUG).

Sybren de Kinderen (VUA) (October 25, 2010).Needs-Driven Service Bundling in a Multi-SupplierSetting. The computational e3-service approach.Vrije Universiteit Amsterdam. Promotor: Prof.dr.J.M. Akkermans (VUA). Copromotor: dr. J. Gordijn(VUA).

Alin Gavril Chitu (TUD) (November 2, 2010).Towards Robust Visual Speech Recognition. DelftUniversity of Technology. Promotores: Prof.dr.C.M. Jonker (TUD), Prof.drs.dr. L.J.M. Rothkrantz(TUD).

Marcel Heerink (UvA) (November 3, 2010).Asessing Acceptance of Assistive Social Robots byAging Adults. University of Amsterdam. Promotor:Prof.dr. B.J. Wielinga (UvA). Copromotores: Dr.ir.B.J.A. Krose, Dr. V. Evers (UvA).

B. Thomee (UL) (November 4, 2010). A Picture isWorth a Thousand Words. Leiden University.Promotor: Prof.dr. J.N. Kok (UL).

Milan Lovric (EUR) (November 5, 2010).Behavioral Finance and Agent-Based ArtificialMarkets. Erasmus University Rotterdam.

Promotores: Prof.dr. J. Spronk (EUR), Prof.dr.ir.U. Kaymak (EUR, TU/e).

Siska Fitrianie (TUD) (November 8, 2010).Human Handheld-Device Interaction: Anadaptive user interface. Delft University ofTechnology. Promotores: Prof.dr. H. Koppelaar(TUD), Prof.drs.dr. L.J.M. Rothkrantz (TUD).

Chen Li (UT) (November 11, 2010). MiningProcess Model Variants: Challenges,Techniques, Examples. Twente University.Promotor: Prof.dr. R.J. Wieringa (UT). Co-promotor: Dr. A. Wombacher (UT).

J.S. de Bruin (UL) (November 23, 2010).Service-Oriented Discovery of Knowledge.Leiden University. Promotor: Prof.dr. J.N. Kok(UL).

Bouke Huurnink (UvA) (November 26, 2010)Search in Audiovisual Broadcast Archives.University of Amsterdam. Promotores: Prof.dr.M. de Rijke (UvA), Prof.dr.ir A.W.M.Smeulders (UvA).

Alia Khairia Amin (CWI) (December 8, 2010).Understanding and Supporting InformationSeeking Tasks in Multiple Sources. EindhovenUniversity of Technology. Promotor: Prof.dr. L.Hardman (CWI/TU/e). Copromotor: Dr. J.R. vanOssenbruggen (VU/CWI).

Peter-Paul van Maanen (VU) (December 9,2010). Adaptive Support for Human-ComputerTeams: Exploring the use of cognitive models oftrust and attention. Vrije UniversiteitAmsterdam. Promotor: Prof.dr. J. Treur (VU).Copromotor: Dr. T. Bosse (VU).

Edgar Meij (UVA) (December 10, 2010).Combining Concepts and Language Models forInformation Access. University of Amsterdam.Promotor: Prof.dr. M. de Rijke (UvA).

Jahn-Takeshi Saito (UM) (December 15,2010). Solving Difficult Game Positions.Maastricht University. Promotor: Prof.dr. G.Weiss (UM). Co-promotores: Dr.ir. J.W.H.M.Uiterwijk (UM), Dr. M.H.M. Winands (UM).

Vincent Pijpers (VU) (December 17, 2010). e3-Alignment: Exploring Inter-OrganizationalBusiness-ICT Alignment. Vrije UniversiteitAmsterdam. Promotor: Prof.dr. J.M. Akkermans(VU). Copromotor: Dr. J. Gordijn (VU).


Section EditorRichard Starmans

SIKS Basic Course“Research Methods andMethodology for IKS”

INTRODUCTIONOn November 24, 25, and 26, 2010, the School forInformation and Knowledge Systems (SIKS)organizes the annual three-day course “ResearchMethods and Methodology for IKS”.

The location will be Conference centerWoudschoten in Zeist.

The course will be given in English and is part ofthe educational Program for SIKS-Ph.D. students.Although the course is primarily intended for SIKS-Ph.D. students, other participants are not excluded.However, their number of passes will be restrictedand depends on the number of SIKS-Ph.D. studentstaking the course.

“Research Methods and Methodology for IKS” isrelevant for all SIKS-Ph.D. students (whetherworking in computer science or in informationscience). The primary goal of this hands-on courseis to enable these Ph.D. students to make a goodresearch design for their own research project. Tothis end, it provides an interactive training invarious elements of research design, such as theconceptual design and the research planning. Butthe course also contains a general introduction tothe philosophy of science (and particularly to thephilosophy of mathematics, computer science andAI). And, it addresses such divergent topics as “thecase-study method”, “elementary researchmethodology for the empirical sciences” and“empirical methods for computer science”.

“Research Methods and Methodology for IKS” isan intense and interactive course. First, all studentsenrolling for this course are asked to read some pre-course reading material, comprising some papers

that address key problems in IKS-methodology.These papers will be sent to the participantsimmediately after registration. Secondly, allparticipants are expected to give a briefcharacterization of their own research project/proposal, by answering a set of questions,formulated by the course directors, and based onthe aforementioned literature. We believe thatthis approach results in a more efficient andeffective course; it will help you to prepareyourself for the course and this will increase thevalue that you will get from it.

COURSE COORDINATORS• Hans Weigand (UvT)• Roel Wieringa (UT)• John-Jules Meyer (UU)• Hans Akkermans (VU)• Richard Starmans (UU)

PROGRAMMore details on the program will be madeavailable in due course.

REGISTRATIONIn the conference center there is a limitednumber of places and there is interest from othergroups in the topic as well. Therefore, an earlyregistration is required. For registration you arekindly requested to fill in the registration form atthe SIKS website.

Arrangement 1 includes single room, all meals,and course material.

Arrangement 2 includes only lunch, dinner andcourse material. So no stay in the hotel and nobreakfast.

Deadline for registration for SIKS-Ph.D.students: November 1, 2010.

After that date, applications to participate will behonoured in a first-come first-serve manner. Ofcourse, applications to participate from otherinterested groups are welcome already. They willreceive a notification whether they canparticipate as soon as possible.

Information for non-SIKS-Ph.D. students:SIKS needs a confirmation from yoursupervisor/office that they agree with thearrangement and paying conditions.


SIKS Basic Courses“Mathematical Methods for IKS” and

“Knowledge Modeling”

INTRODUCTIONOn December 7-10, 2010, the School forInformation and Knowledge Systems (SIKS)organizes two basic courses “MathematicalMethods for IKS” and “Knowledge Modeling”.Both courses will be given in English and are partof the obligatory Basic Course Program for SIKS-Ph.D. students. Although these courses areprimarily intended for SIKS-Ph.D. students, otherparticipants are not excluded. However, theirnumber of passes will be restricted and depends onthe number of SIKS-Ph.D. students taking thecourse.

Location: Landgoed Huize Bergen, VughtDate: December 7-10, 2010

SCIENTIFIC DIRECTORS• Prof.dr. Eric Postma (UvT), Mathematical

Methods for IKS• Prof.dr. Tom Heskes (RUN), Mathematical

Methods for IKS• Dr. Bert Bredeweg (UVA), Knowledge

Modeling

PROGRAMThe program is not available yet, but may includethe following topics:

Mathematical Methods for IKS• Basic formalisms relevant to modern intelligent

knowledge systems• Automatically acquired knowledge

representations• Inductive Learning• Bayesian Statistics• Entropy and Information Theory• Computer intensive techniques• Minimum Description Length Methods

Knowledge Modeling• Ontologies, Epistemology and Models• Modeling with Description Logics• Methodology for Ontology Engineering• KADS• OWL: Ontology Language for the Web• Ontology Patterns, Re-use of information

REGISTRATIONIn the conference center there is a limitednumber of places and there is interest from othergroups in the topic as well. Therefore, an earlyregistration is required.

Deadline for registration for SIKS-Ph.D.students: November 30, 2010. After that date,applications to participate will be honoured in afirst-come first-serve manner. Of course,applications to participate from other interestedgroups are welcome already. They will receive anotification whether they can participate as soonas possible.

For registration you are kindly requested to fill inthe registration form at the SIKS website.

Arrangement 1 includes single room, all meals,and course material.

Arrangement 2 includes two lunches, one dinnerand course material. So no stay in the hotel andno breakfast.

Advertisements in theBNVKI Newsletter

Do you want to place a (job) advertisement in theNewsletter of the BNVKI?

• Whole page: € 400 for 1 issue; € 600 for2 subsequent issues; € 900 for 6 subsequentissues.

• Half page: € 300 for 1 issue; € 450 for2 subsequent issues; € 675 for 6 subsequentissues.

You reach an audience of AI professionals,academics and students. Your logo (with link toyour company) will also be shown on theBNVKI/AIABN website during the period ofadvertisement.

Contact [email protected] foradditional information.



BOARD MEMBERS BNVKIProf.dr. A. van den Bosch (chair)Universiteit van Tilburg, Faculteit der LetterenTaal en InformaticaP.O. Box 90153, 5000 LE TilburgTel.: + 31 13 4663117. E-mail: [email protected]

Prof.dr. A. Nowé (secretary)Vrije Universiteit Brussel, Computational Modeling LabDepartment of Computer SciencePleinlaan 2, B-1050 Brussels, BelgiumTel.: + 32 2 6293861. E-mail: [email protected]

Dr. M.V. Dignum (treasurer and vice-chair)Delft University of TechnologyDept. Technology, Policy and ManagementSection Information and Communication TechnologyP.O. Box 5015, 2600 GA DelftTel.: + 31 15 2788064. E-mail: [email protected]

Dr. J.W.H.M. Uiterwijk (BNVKI Newsletter)Universiteit MaastrichtDepartment of Knowledge Engineering (DKE)P.O. Box 616, 6200 MD MaastrichtTel: + 31 43 3883490. E-mail: [email protected]

Prof.dr. M.F. Moens (PR and sponsoring)KU Leuven, Departement ComputerwetenschappenCelestijnenlaan 200A, 3001 Heverlee, BelgiumTel.: + 32 16 325383. E-mail: [email protected]

Dr. A. ten Teije (student affairs)Vrije Universiteit AmsterdamDept. of AI, Knowledge Representation and Reasoning GroupRoom T343, De Boelelaan 1081A, 1081 HV AmsterdamTel.: + 31 20 5987721. E-mail: [email protected]

Dr. K. Hindriks (AI & Industry)Delft University of TechnologyMediamatica Department, Man-Machine Interaction GroupRoom HB12.050Mekelweg 4, 2628 CD DelftTel.: + 31 15 2782523. E-mail: [email protected]

Dr R. Booth (international affairs)University of LuxembourgFaculty of Science, Technology and Communication Room E1016 rue Coudenhove Kalergi, L-1359 LuxembourgTel.: + 352 621 402 011. E-mail: [email protected]

EDITORS BNVKI NEWSLETTERDr. J.W.H.M. Uiterwijk (editor-in-chief)See above for address details

Prof.dr. E.O. PostmaTilburg UniversityFaculty of Humanities, TiCCP.O. Box 90153, 5000 LE TilburgTel: + 31 13 4662433. E-mail: [email protected]

Prof.dr. H.J. van den HerikTilburg UniversityFaculty of Humanities, TiCCP.O. Box 90153, 5000 LE TilburgTel.: + 31 13 4668118. E-mail: [email protected]

M. van Otterlo, M.Sc.University of Twente, Dept. of Computer ScienceP.O. Box 217, 7500 AE EnschedeTel.: + 31 53 4894111. E-mail: [email protected]

Dr. L. Mommers (section editor)Universiteit Leiden, Dept. of Meta-JuridicaP.O. Box 9520, 2300 RA LeidenTel.: +31 71 5277849. E-mail: [email protected]

J. De Beule, M.Sc. (editor Belgium)Vrije Universiteit Brussel, Artificial Intelligence LaboratoryPleinlaan 2, B-1050 Brussels, BelgiumTel.: +32 2 6293703. E-mail: [email protected]

Dr. R.J.C.M. Starmans (section editor)Manager Research school SIKS,P.O. Box 80089. 3508 TB UtrechtTel.: + 31 30 2534083/1454. E-mail: [email protected]

Dr. K. Hindriks (section editor)See above for address details

HOW TO SUBSCRIBEThe BNVKI-AIABN Newsletter is a direct benefit ofmembership of the BNVKI-AIABN: Benelux Association forArtificial Intelligence. Membership dues are€ 40 for regular members; € 25 for doctoral students (AIO’s);and € 20 for students. In addition, members will receive access tothe electronic version of the European journal AICommunications. The Newsletter appears bimonthly andcontains information about conferences, research projects, jobopportunities, funding opportunities, etc., provided enoughinformation is supplied. Therefore, all members are encouragedto send news and items they consider worthwhile to the editorialoffice of the BNVKI/AIABN Newsletter. Subscription is doneby payment of the membership due to Postbank no. 3102697 inThe Netherlands (IBAN: NL 74 PSTB 0003 1026 97; BIC:PSTBNL21). Specify BNVKI/AIABN as the recipient, and pleasedo not forget to mention your name and address. Sending of theBNVKI/AIABN Newsletter will only commence after yourpayment has been received. If you wish to conclude yourmembership, please send a written notification to the editorialoffice before December 1, 2010.

COPYThe editorial board welcomes product announcements, bookreviews, product reviews, overviews of AI education, AIresearch in business, and interviews. Contributions statingcontroversial opinions or otherwise stimulating discussions arehighly encouraged. Please send your submission by E-mail (MSWord or text) to [email protected].

ADVERTISINGIt is possible to have your advertisement included in theBNVKI/AIABN Newsletter. For further information aboutpricing etc., see elsewhere in the Newsletter or contact theeditorial office.

CHANGE OF ADDRESSThe BNVKI/AIABN Newsletter is sent from Maastricht. TheBNVKI/AIABN board has decided that the BNVKI/AIABNmembership administration takes place at the editorial office ofthe Newsletter. Therefore, please send address changes to:

Editorial Office BNVKI/AIABN NewsletterMarijke VerheijDepartment of Knowledge Engineering (DKE),Maastricht UniversityP.O. Box 616, 6200 MD Maastricht, the NetherlandsE-mail: [email protected]://www.bnvki.org

















http://www.bnvki.org

newsletterii.tudelft.nl/bnvki/wp-content/uploads/2017/03/27.5.pdf · there is a famous conjecture,...

Documents