An Emotional-Evolutionary Te hnique forLow-Level Goal De�nition in a Multi-PurposeArti� ial CreaturePatrí ia de Toro1, Ri ardo Gudwin1, and Mauro Miskulin2

1 DCA/FEEC/UNICAMPP.O. Box 6101 Campinas,Brazil, 13083-970{patbgi,gudwin}�d a.fee.uni amp.br2 DSIF/FEEC/UNICAMPP.O. Box 6101, Campinas-SP,Brazil, 13083-970mauromiskulin�terra. om.brAbstra t. Arti� ial Creatures are embodied autonomous agents livingin a virtual world, like e.g. in a omputer game or in ethologi al simula-tion studies. Usually, these arti� ial reatures may have general high-levelgoals, like survival, killing opponents, feeding, mating, dis overying theworld, et . These general high-level goals must be turned into low-levelgoals in time and spa e in order for the agent to generate its �nal behav-ior within environment. This paper presents an emotional-evolutionaryte hnique whi h en odes ea h high-level goal into a parti ular emotion,and using a blend of su h emotions as a �tness fun tion in a geneti algorithm, evolves parti ular goals in time and spa e in order to drivethe reature's behavior.1 Introdu tionThe resear h �eld of autonomous intelligent agents omprehends a powerful ab-stra tion for many kinds of pra ti al appli ations, from mobile roboti s to om-puter games. One of the lassi al problems in this ontext is the problem ofautonomous navigation in omplex environments. This problem appears e.g. inthe ase of mobile roboti s, where a mobile robot needs to de ide a traje tory,sin e an initial point up to a target, without olliding with obsta les (possiblyminimizing the distan e overed and/or the time of travelling). Another exam-ple of the same problem would be the development of intelligent opponents in omputer games, where an intelligent ontrol system must de ide the a tions tobe provided by an agent in order to foster a good entertainment to the systemuser, simulating with realism the behavior of a human opponent.Several alternative frameworks are possible in order to deal with this problem,depending on the aspe ts we want to emphasize. One possible form to systemizethe problem is to onsider it in the sense of an arti� ial reature whi h exists ina ertain environment, moving itself in this environment and a ting on it [2℄.

Published in Angelo Loula & João Queiroz (Eds), Advances in Modeling Adaptive and Cognitive Systems. UEFS, 2010.

48

Classi al solutions for this problem are well known in the �eld of arti� ialintelligen e [17℄ usually involving state ma hines, sear h algorithms and, even-tually, logi programming. More advan ed algorithms may use neural networks,fuzzy logi or evolutionary omputation.More re ently, the on ept of emotion, as brought from ognitive psy hologyand philosophy, was suggested in the literature, as an alternative way of dealingwith this problem [3, 4, 16, 15, 22, 7, 8, 23, 18, 19, 6, 13℄.There is no onsensus, though, on what exa tly are emotions. Di�erent ap-proa hes have di�erent views for what it is and how to model them. For example,Ortony [14℄ views emotions as �valen ed rea tions to events, agents, or obje ts,with their parti ular nature being determined by the way in whi h the eli itingsituation is onstrued�. Sloman [20, 21℄ understand emotions as internal �alarms�whi h give a momentary emphasis to ertain groups of signals. Damasio [9, 10℄make a distin tion between �emotions�, whi h a�e t the body and �feelings�,whi h are a ognitive introspe tion of an emotion. Other authors may have fur-ther di�erent views for what emotions are. For some of them, emotions work like�ampli�ers� for motivations. For others they are homeostati pro esses relatedto physiologi al variables [7℄. Some authors, instead of a single on ept of emo-tion, develop a omplex �emotional system�, where many di�erent on epts like�motivations�, �drives�, �impulses�, �a�e tions�, �needs� and other terms are usedto represent di�erent aspe ts of this emotional system.In arti� ial reatures [4℄, emotions are usually employed in order to drivebehavior, being used as a riteria for a tion-sele tion me hanisms. A problem,though remains to be solved. In a multi-purpose arti� ial reature (e.g. a har-a ter in a omputer game), there are many di�erent high-level goals whi h needto be satis�ed, like survival, killing opponents, feeding, mating, dis overying theworld, et . The reature needs to de ide, at ea h point in time and spa e, what todo next, and so de ide the next lower-level goals, like where to move to, a tionsto be performed, et .In this arti le, mixing ideas from emotion-based ontrol systems and evolu-tionary omputation, we present a te hnique for high level goal de�nition in the ase of a multi-purpose arti� ial reature.The ar hite ture in whi h this work is based on, was suggested originallyin [11℄, whi h de�nes an autonomous reature equiped with a sensorial andmotor apparatus apable to navigate through an environment full of obje tswith di�erent hara teristi s. Obje ts do have � olors", su h that ea h � olor� isasso iated to: a degree of �hardness� (varying from 0 to 1), whi h works like afri tion oe� ient in order to slow down the reature movement (or totally blo kit) over pla es where there are obje ts of this olor; a �taste�, whi h an be bador good (varying from -1 to 1); and an ability to drain or supply �energy� (whi halso may vary from -1 to 1). The reature owns an internal battery whi h isre hargeable. Tou hing an obje t whi h supplies or drains energy would hangethe harge of this battery. The reature an navigate in this environment, andmany goals may be pres ribed. An elementary goal will be to navigate from aninitial point up to a target point, avoiding ollision with some undesirable obje ts49

PSfrag repla ementsTargetDe ision-MakingThiswork High-level Control

(x, y) (x, y) Point to be rea hed (target)Dire t ControlVirtualEnvironmentSensors A tuators Ar hite ture asproposed by[11℄Fig. 1. Arti� ial Creature Ar hite tureat the environment. There may be, however, some ases in whi h this ollision isdesirable (as in the ase of obje ts that work as energy suppliers). The generalgoal would be to generate the agent behavior a ording to the di�erent purposesthat may be attributed to it.Two levels of ontrol have been onsidered: a lower level of dire t ontrol anda hierar hi ally superior level, of target de ision-making (see Figure 1). At thelevel of dire t ontrol, the input is a target point to be rea hed by the reature.The reature must move from its urrent position up to this target point, without ollisions with obsta les. At the level of target de ision-making, the system mustde�ne the target points to where the reature must move itself, onsidering a setof high level general purposes for the reature's existen e. Together, the levels ofdire t ontrol and target de ision-making generate a omplex behavior for the reature. The level of dire t ontrol was extensively addressed in [11℄, by meansof a rea tive/deliberative strategy whi h is responsible for the generation andexe ution of plans for the navigation to the target point. In this work, we extendthe work of [11℄ by developing the superior level of target-de ision-making.The level of dire t ontrol only deals with the low-level goal of safe movingthe reature along a set of obsta les. The level of target de ision-making, allowsfor the onsideration of a high levels set of general purposes for the existen eof the reature. Examples of high level purposes in our ase are: to explorethe environment, taking are of the reature's energy balan e and to learn anoptimal poli y regarding the distan e to maintain to obsta les in order to havea safe navigation. To generate a behavior whi h onsiders together all thosepurposes, we made use of three emotional metaphors: uriosity, hunger and fear.Curiosity will be responsible for the reature's exploratory behavior. Hunger willbe responsible for the maintenan e of the reature's energy level. And fear must�ne-tune the ollision avoidan e behavior.50

Fig. 2. Sensory-Motor Stru ture of the Creature2 Experiments and pro eduresA sket h of the reature's sensory-motor stru ture an be seen at Figure 2. The reature does not have an �a priori" map of the environment. Based on the sen-sorial information, the reature builds an in remental map of the environment,whi h is used to generate movement plans. A more detailed a ount of the rea-ture stru ture and dynami model used in the simulator and the internals of thelow level dire t ontrol an be found in [11℄.In this work, extending the work in [11℄, we deal with the high level targetde ision-making sub-system, whi h de�nes at ea h moment, the next point towhere the reature should move. A ording to [2℄, an arti� ial reature mayhave four types of standard behavior: (i) appetitive: behavior is dire ted towardsan attra tive obje t or situation; (ii) aversive: behavior is dire ted away fromnegative situations; (iii) exploratory: behavior is dire ted toward stimuli that arenovel in the environment and, �nally, (iv) neutral behaviors relating to obje tsthat are neither appetitive nor aversive. It is interesting to note that our hoi eof �hunger�, �fear� and � uriosity� orresponds to the hoi e of an appetitive, anaversive and an exploratory behavior. In this experiment, we do not ontemplatethe �neutral� behavior.The level of dire t ontrol (as in [11℄), already in orporates the emotion �fear"during the generation of a plan. This plan only onsiders the obje ts it alreadyknows to exist at the environment. However, in [11℄ this fear was de�nidedby heuristi means, en oded by a given (hard- oded) utility fun tion. Here, wedeveloped a more �exible me hanism, where this fear an in rease or diminish,as su h as the reature enters in onta t with obje ts that represent obsta les.Or, in other words, the reature learns to have �fear" while intera ting ( olliding)to unpleasant obje ts. Moreover, this fear is also used at the determination ofthe reature's next target. The determination of the next target is not a simple51

task. Depending on the situation, di�erent behaviors an be desirable. In the asewhi h the reature's energy level is high, we would expe t the reature to developan exploratory behavior, going to parts of the environment whi h are not alreadyknown. However, in the ase whi h the energy level is low, we would expe t the reature to sear h within its map for a known power sour e and go to there.After being re-energized, the reature may return to the exploratory behavior. Toimplement this hybrid behavior, we developed two distin t strategies, one whi his hard- oded, expli itly reating some rules of priorities among the emotions,and another one that makes an automati blend of all emotions and derives a�nal de ision, based on a geneti algorithm.2.1 Des ription of the Hard- oded AlgorithmIn our �rst experiment, emotions are onsidered in a hard- oded way, givingexpli ity priority on some emotions before others.In this algorithm, if the energy level is higher than a ertain level, a randompoint is hosen to be the next target. If the uriosity fun tion asso iated to thispoint is below a ertain level, than another random point is generated, until aquali�ed point is rea hed. The algorithm uses then the dire t ontrol de�ned in[11℄ to develop a plan whi h the reature puts into exe ution until the targetis met. This step implements only impli itly the emotions of fear and uriosity.Fear is just a metaphor for the utility fun tion whi h is used to derive theplan by dire t ontrol. Curiosity is also just a metaphor for the utility fun tionwhi h omputes the degree of "awareness" of a given point. This is the standardbehavior for the reature.However, in the ase whi h the energy level is below a ertain limit, a newbehavior starts to run. The reature stops seeking its standard target, saving itto resume in the future, and start looking for the nearest point (in its map) whereit knows there is an energy supplier lo ated. It uses than the dire t ontrol tobuild a plan to rea h this point. After getting in onta t with the energy supplier,the reature waits until its energy level in reases up to the maximum, and afterthat, it restores the original target ba k, with a new plan given by the dire t ontrol.It may be observed that the hard- oded algorithm imposes expli itly a �xedway of intera tion among the emotions, leading to a �xed set of behaviors, a - ording to the following rules:� Fear (to ollide with obsta les): the algorithm normally sear h to prevent ollisions. If a ollision happens, the agent in reases its � aution fa tor�;� Curiosity (sear h for unknown points at the environment): Curiosity mea-sures whether a given point is already known by the agent;� Hunger (to keep energy balan e): Hunger leads to an ex eption behavior,where uriosity is dis onsidered in order to generate the next target.The hard- oded algorithm has many limitations. Perhaps, the most obviousone is that the emotional metaphor, although useful in order to inspire thealgorithm development, does not allow for a generalized way to deal with di�erent52

PSfrag repla ementsx1

x2

x3

y1

y2

y3

{

x3 = αx1 + (1 − α)x2

y3 = αy1 + (1 − α) y2Population Size = 1000 individualsCrossover Prob.= 0,5Mutation Prob.= 0.1Fig. 3. Geneti Algorithm Parameters and Operatorssituations. The intera tion among emotions need to be treated ase by ase.When it is desirable to in lude new emotions within the system, these newemotions shall require an extended spe i� ation on how it will intera t with allthe other already implemented emotions. The treatment of on�i ting emotionsmay require the generation of ex eptional behavior, whi h may require a spe ialtreatment nearly ase by ase. To solve these restri tions, we tried to develop analgorithm whi h ould be generalized in an uniform way, not requiring spe ialbehaviors, and allowing for the addition of new emotions without too mu he�ort. The result is the following algorithm.2.2 Des ription of the Geneti AlgorithmIn order to solve the aforementioned problems, a geneti algorithm (Figure 4)was on eived, where a population of possible target points shall evolve, using asa �tness fun tion a ombination of all the utility fun tions, one for ea h designedemotion. This ombination results in a fun tion of �desirability�, that will mixfear, hunger and uriosity.Initially this population of �potential targets� is generated randomly. At ea hevolutive step, these �potential targets� start to evolve, be oming on�ned tosome regions of spa e where the �desirability� is higher. For the geneti algorithm,ea h target is en oded in the form of a hromosome ontaining a position (x,y)in the two-dimensional spa e. The rossover between two hromosomes is madeby means of a linear ombination between the two points represented by ea h hromosome, resulting in a point that is lo ated somewhere in the straight linewhi h bounds the two points (see Figure 3). A mutation operator translates agiven point somewhere into its near neighborhood. All the possible goals areGenerate

Initial PopulationEvaluate

Population

Termination Condition

met ?

Obtain BestIndividual

CrossoverMutationSelection

Start

ResultYes

NoFig. 4. The Geneti Algorithm53

Fig. 5. A Plain View of the Environ-ment Fig. 6. Representation of the Emo-tional Fa tor Curiosity (di�erent levelsof gray). Green spots are the possibletargetsFig. 7. Optimized Situation: reaturewith a higher knowledge of the environ-ment Fig. 8. Situation where the energy levelis low: Targets are on entrated aroundenergy suppliersevaluated regarding its �desirability�, and a pro ess of elitist sele tion only keepsin the population the potential targets with greater �desirability�.Figures 5, 6 and 7 present the details for an example of simulation using thegeneti algorithm. Figure 5 shows the environment where the reature developsits navigation. Figures 6 and 7 represent the internal map onstru ted by the reature, in di�erent instants of the simulation. Figure 6 shows a situation in thebeginning of the simulation, where the dark areas represent unknown parts ofthe environment, implying in great uriosity and so great desirability for possibletargets. Please pay attention to the distribution of light-green points inside theseareas. They represent the population of possible targets being evolved by thegeneti algorithm. Figure 7 shows a further time in simulation, where the reaturealready had the han e to know other areas. Observe that in this ase, thenumber of dark areas is lower, and the possible targets are on entrated in asmaller range of lo ations.Figure 8 shows another simulation (in a di�erent environment), in a situationwhile the reatures's energy resour es are too low. Noti e that, in this ase,

54

the lo ations where there are obje ts with energy supply do attra t a biggerdesirability. Pay attention to the on entration of possible targets under theseobje ts.3 Simulation ResultsIn order to evaluate the simulation results, we generated some measurings, whi hwe present in the following. Figures 9 and 10 show the variability of the reature'senergy level over time. Figure 9 shows the results for the hard- oded algorithmand Figures 10 shows the result for the geneti algorithms with populationsof 1000, 3000 and 5000 individuals. Observe the behavior in both ases. The reature starts will full power and have the energy de reasing until it rea hesthe level of around 30 or 40 % when it then get refueled. The behavior is moreor less the same both for the hard- oded algorithm and the geneti algorithm.Figures 11 and 12 show the variability of the reature's fear over time. Fear is omputed as a fun tion whi h al ulates the distan e to the losest undesirableobje t at ea h time. We ould not get any kind of orrelation between thesegraphi s, showing that ompletely di�erent solutions where performed by ea h ase of the algorithms.Figures 13 and 14 show the variability of the unknown parts of the worldover time. Figure 13 shows the ase for the hard- oded algorithm and Figure 14shows the ase for the geneti algorithms, with populations of possible targetsof 1000, 3000 and 5000 individuals.We an see that in both ases, the reature in reases its knowledge of theenvironment up to 100% after some time. In this ase, we dete ted a smalldi�eren e in e� ien y among the many algorithms we simulated. The hard- oded algorithms is less e� ient, gaining total knowledge of the environment onlyaround 430s. The geneti algorithm with population size of 5000 individuals getthe same 100% with around 380s. The geneti algorithm with population sizes of3000 and 1000 individuals also performed better than the hard- oded algorithm

Fig. 9. Variation of Creature's EnergyLevel over time: Hard- oded Algorithm Fig. 10.Variation of Creature's EnergyLevel over time: Geneti Algorithm55

Fig. 11. Variation of Creature's Fearover time: Hard- oded Algorithm Fig. 12. Variation of Creature's Fearover time: Geneti Algorithm

Fig. 13. Variation of Unknown Partsof the Environment over time: Hard- oded Algorithm Fig. 14. Variation of Unknown Parts ofthe Environment over time: Geneti Al-gorithmAs a general evaluation, we may on lude that both the hard- oded and thegeneti algorithm do have a similar behavior, a hieving their main generi goalsof exploring the environment, maintaining their energy level and avoiding olli-sions with obsta les, with a slight di�eren e to the geneti algorithm. But, onthe other side, the geneti algorithm has a great advantage in terms of s ala-bility. For the hard- oded algorithm, for ea h new emotion that we would liketo in lude, all the algorithm needs to be redesigned. In the ase of the geneti algorithm, only a small hange in the desirability fun tion is required, if we in-tend to in lude a newer emotion. As a on lusion, we may rea h that the geneti algorithm is more advantageous than the hard- oded one.4 Dis ussionThe problem of a tion sele tion is treated in many di�erent ways in the literature.During the 90's, the older entralized Planning-Exe ution (Hierar hi al) Meth-ods, as e.g. in [1℄ were put forward in favor of distributed, behavior based, using56

spreading a tivation te hniques like e.g. in [4℄ or [12℄. Many di�erent approa hes,like Brook's subsumption ar hite ture [5℄, Blumberg's motivation based a tionsele tion [4℄ and Maes' Behavioral Networks [12℄ were proposed as alternativesto the problem. Even though these distributed behavior based methods wereable to perform very sophisti ated behavior, they never rea hed the level of so-phisti ation of planning ahead many steps of a tivity. In other words, to havein advan e a plan for the future is a typi al human strategy whi h an not beover omed totally by means of oportunisti behavior-based a tions. On the otherside, to try planning everything up to the last details is learly unrealisti , asthe world is onstantly hanging, and so plans should be amenable to orre -tions or at least re-planning in order to deal with this un ertainty. The urrentapproa h, developped in this work, is in some sense a kind of a hybrid betweenold-fashioned plan-based approa hes and non-deterministi behavior-based nou-velle AI algorithms. The inspiration for the emotional-evolutionary te hniquedevelopped here is a tually based on how human beings de ide their ourse ofa tion in daily a tivity. We usually have general goals, most of them geneti allydetermined, like foraging, mating, a umulating resour es (in luding goods andknowledge), having fun, whi h are enfor ed in the evaluation of our a tions byour �value system�. This value system manifests itself by means of our emotions,whi h are used to lassify the many �ideas� of possible a tions we have duringour daily life. In this sense, a population of �ideas� of possible a tions ompeteamongst ea h other (should I go to the movies or should I exer ize myself at theGym ?), and we use our emotions to de ide what to do next. So, we on eivedan evolution of a population of ideas, ompeting to ea h other based on theemotional response they are able to arise. Ea h emotion, on the other side, isnothing more than an en oding of a high-level purpose (the many purposes wehave for our life), whi h are used to drive our daily de isions. This high-levelemotional-evolutionary algorithm ould be viewed as the behavior-based part ofour approa h. As soon as we de ide what to do, our low-level sub-system takes are of everything, and reates a plan whi h will lead us to the a hievement ofour idea. This is the old-fashioned hierar hi al planning system, assuming the ontrol until something unexpe ted o urs and everything needs to be re-doneagain. This is how we live, mixing some oportunisti emotionally based de isionswith some planning guiding our daily life. This is not a �nal algorithm though.During our daily life, we have di�erent time-s ale plans, reated to attend long-term goals whi h require a long-term ommitment to some a tivity. We makeplans for our very next steps, like passing through the door and going to the o�ema hine, we plan for where to go at the weekend, and we also plan our job, our areer, marriage and other long-term plans. These short, middle and long-termplans usually have to live with ea h other, sometimes with on�i ts betweenthem. In order to attend the long-term plans, we usually engage ourselves in aroutine s hedulle of a tivities like going to work, having lun h at a ertain time,having dinner, oming ba k home and going to sleep at pre-determined times.In the between, we have to de ide when to go to the restroom, or to make apause in our work and have a onversation to a olleague, or to have a so ial or57

family entertainment. We are onstantly managing many di�erent goals, some ofthem short-term, other long-term goals, and always making a trade-o� betweenthe oportunism of some a tion and the routine of a long-term plan or s hedulle.Our urrent algorithm doesn't onsider this mixing of short term and long termgoals, even though gives some lues on how to deal with them. This should beaddressed in a future work.5 Con lusionsIn this paper, we presented an emotional-evolutionary te hnique for low levelgoal de�nition in a multi-purpose arti� ial reature. This te hnique en odesea h high level goal (like surviving, maintaining energy balan e or in reasingknowledge) into a separate emotion (fear, hunger, uriosity), and then a geneti algorithm is used to evolve a population of andidate low-level goals (like runningaway from some danger, eating something or going to an unknown part of theenvironment), in order to drive the �nal reature behavior. This strategy, up toa ertain point, would be omparable to how low-level human goals are hosenamongst a population of andidate ideas. Ea h idea is evaluated by means of theemotional value it is able to arouse, and the most prominent idea, in emotionalterms, is hosen to be performed. There is a random omponent, brought bygeneti algorithm, and also some kind of intelligent behavior, as the emotions arehigh-level goals en oded in a normalized form. We understand that the urrentexperiments only suggest this insight, being still not de�nitive. But the presentstudy shows the viability of su h te hnique, and motivates us to ontinue thisline of investigation.6 A knowledgmentsPatrí ia de Toro is thankful to the CAPES for the �nan ial support in the form offellowship. Ri ardo Gudwin would like to thank the Brazilian National Resear hCoun il (CNPq) for grant #304649/2004-0Referen es1. J. Albus. An outline for a theory of intelligen e. IEEE Transa tion on Systems,Man and Cyberneti s, 21(3):473�509, May/June 1991.2. C. Balkenius. Natural Intelligen e in Arti� ial Creatures. Lund Univ. CognitiveStudies 37, 1995.3. J. Bates. The role of emotion in believable agents. Communi ations of the ACM,37(7):122�125, 1994.4. B. M. Blumberg. Old Tri ks, New Dogs: Ethology and Intera tive Creatures. PhDthesis, MIT Media Lab, Cambridge, MA, 1996.5. R. A. Brooks. New approa hes to roboti s. S ien e, 253:1227�1232, September1991.6. D. Budakova and L. Dakovski. Computer Model of Emotional Agents, volume 4133.Springer Berlin / Heidelberg, 2006.58

7. L. Cañamero. A hormonal model of emotions for behavior ontrol. In Pro . of the4th ECAL (ECAL'97), 1997.8. L. Cañamero. Issues in the design of emotional agents. In A. Press, editor, In:Pro . of Emotional and Intelligent: The Tangled Knot of Cognition. AAAI FallSymposium, pages 49�54, Menlo Park, CA, 1998.9. A. R. Damásio. O Erro de Des artes: Emoção, Razão e Cérebro Humano. Europa-Améri a, 1994.10. A. R. Damásio. The Feeling of What Happens: Body and Emotion in the Makingof Cons iousness. New York: Har ourt, 1999.11. R. R. Gudwin. Contribuições ao Estudo Matemáti o de Sistemas Inteligentes. PhDthesis, DCA-FEEC-UNICAMP, 1996.12. P. Maes. How to do the right thing. Conne tion S ien e Journal, 1(3), 1989.13. J. J. C. Meyer. Reasoning about emotional agents. Int. J. Intell. Syst, 21(6):601�619, 2006.14. A. Ortony, G. Clore, and A. Collins. The Cognitive Stru ture of Emotions. Cam-bridge Univ. Press, 1998.15. R. W. Pi ard. A�e tive Computing. MIT Press, Cambridge, 1997.16. W. S. N. Reilly. Believable So ial and Emotional Agents. PhD thesis, CarnegieMellon University, Pittsburgh, PA, 1996.17. S. J. Russell and P. Norvig. Arti� ial Intelligen e: A Modern Approa h. Prenti eHall, Englewood Cli�s, NJ, 1995.18. L. M. Sarmento. An emotion-based agent ar hite ture. Master's thesis, FC Uni-versity of Porto, 2004.19. C. Septseault and A. Nédéle . A model of an embodied emotional agent. In IVA,page 498, 2005.20. A. Sloman. Damasio, des artes, alarms, and meta-management. In In Pro . of theIEEE International Conf. on Systems, Man, and Cyberneti s, pages 2652�2657,1998.21. A. Sloman. Beyond shallow models of emotion. In Cognitive Pro essing, volume 2,pages 177�198, 2001.22. J. Velásquez. A omputational framework for emotion-based ontrol. In In Pro . ofthe Grounding Emotions in Adaptive Systems - workshop SAB'98, Zuri h, Switzer-land, 1998.23. R. Ventura. Emotion-based agents. Master's thesis, Inst.Sup.Te .-Lisboa,Portugal,2000.

59