decision-making and rationality - university of...
TRANSCRIPT
2015/10/7
1
Resilience Informatics for Innovation Classical Decision Theory
RERC/TMIKazuo FURUTA
Decision-making and rationality
What is decision-making?n Methodology for making a choicen The quality of decision-making determines success
or failure of innovation and business.
Rational decision-makingn No means exist that certainly lead us to a correct
decision due to uncertainties in the world.n Anyhow, we need to decide rationally in some
sense, with no regret. ⇒ Max. use of information
2015/10/7
2
Classical decision-making
Normative model of decision-makingn Choosing one option from manyn Utility is what one wants to make maximum in the
choice.
Option 1Option 2Option 3
…… Option i
Assess options
Collectinformation
Predefined options
ChoiceUtility of options(u1, u2, u3, ....)
Choice and preference
Preferencen An act to chose one option from manyn Basic unit of classical decision
Order of preference in a dyada is preferred to b.a is equivalent to b.a is preferred or equivalent to b.
ba ≈ba f
ba f
2015/10/7
3
Utility and utility function
Utilityn What one wants to make maximum in the choicen Subjective value of each option evaluated by the
decision-maker
Utility function u(x)n Mapping from the set of options to real numbers
n Utility function is a tool to deal with the utility mathematically
)()( buauba ≥⇔f
Condition for rational decision
In order that the preference is mathematically consistent, it must be a weak order relation.
n Completeness or for any pair of a and b
n Transitivity
ba f ba p
cacbba fff ⇒,
2015/10/7
4
Expected utility
Definition of expected utility
si : possible state pi : occurrence probability of si u(si) : utility of si
Expected utility hypothesisn The utility of option a under uncertainty is
represented by its expected utility E(a).
∑=
=n
iii psuE
1)(
Example of expected utility
Win.prob. Prize ConsolationLottery 1 0.01 $15,000 noneLottery 2 0.02 $8,000 noneLottery 3 0.01 $10,000 $50
E(L1) = 0.01 x $15,000 = $150E(L2) = 0.02 x $8,000 = $160E(L3) = 0.01 x $10,000 + 0.99 x $50 = $149.5
2015/10/7
5
Choice under uncertainty
Uncertain situation si : Possible state pi : Occurrence probability of si
aj : Option of action uij : Utility when si obtains after aj has
been taken
Various decision criteria (1)
Expected utility criterion
n If the occurrence probability distribution is unknown, suppose pi = 1/n (Laplace criterion)
Max-min criterion
n Maximize the utility for the worst case
max)(s.t.*1
→== ∑=
n
iiijjj puaUaa
maxmin)(s.t.* →== ijijj uaUaa
2015/10/7
6
Various decision criteria (2)
Hurwitz criterion
n α = 0 and α = 1 correspond to complete pessimism (max-min criterion) and complete optimism.
Regret criterion (Minimum opportunity loss)
n Opportunity loss rij is the degree of regret compared with the maximum utility obtainable with perfect foresight.
maxmin)1(max)(s.t.* →−+== ijiijijj uuaUaa αα
ijijjij
ijijj
uurraLaa
−=→==
maxminmax)(s.t.*
Example of choice under uncertainty (1)
You are a street food stall owner. Which sells well depends on the weather: ice cream or hot dog. Which will you stock more for tomorrow’s business.
Weather Sunny Cloudy RainyProbability 0.6 0.3 0.1Ice cream 1,000 500 200Half & half 800 700 450Hot dog 650 600 500
2015/10/7
7
Example of choice under uncertainty (2)
Weather Hurwitz* RegretIce cream 600 300Half & half 625 200Hot dog 575 350
Criterion Expected Laplace Max-minIce cream 770 567 200Half & half 735 650 450Hot dog 620 583 500
* α = 0.5
Conditional probability
U : Sample spaceE, F : Subset of U| X | : Size of X
Conditional probability
UE F
)()(
||||||||
||||)|(
FPFEP
UFUFE
FFEFEP ∩
=∩
=∩
=
)|()()|()()( FEPFPEFPEPFEP ==∩
2015/10/7
8
Example of conditional probability
Two invisible bins A and BA : 3 silver & 1 gold coinsB : 2 silver & 4 gold coins
n You drew a silver coin from either of the two bins by chance. Which bin did you choose?
692.0139
6/18/38/3
)()()|( ≈=
+=
∩=
SPSAPSAP
A
start
B
1/2G
G
S1/2
3/4 S
1/4
2/6
4/6
3/8
1/8
1/6
2/6
Independent events
When event E and F satisfy the following conditions, they are independent.
The following is the necessary and sufficient condition for that E and F are independent.
)()|()()|( FPEFPEPFEP ==
)()()( FPEPFEP =∩
2015/10/7
9
Bayes theorem
General form of Bayes theoremn H1, H2, … Hn are exclusive each other and
)()()|()|(
EPFPFEPEFP =
UHHH n =∪∪∪ L21
)|()()|()()|()()|(
11 nn
iii HEPHPHEPHP
HEPHPEHP++
=L
Coins-in-bins problem revisited
Two invisible bins A and BA : 3 silver & 1 gold coinsB : 2 silver & 4 gold coins
n You drew a silver coin from either of the two bins by chance. Which bin did you choose?
139
2/13/12/14/32/14/3
)()|()()|()()|()|( =
⋅+⋅⋅
=+
=BPBSPAPASP
APASPSAP
A
start
B
1/2G
G
S1/2
3/4 S
1/4
2/6
4/6
3/8
1/8
1/6
2/6
2015/10/7
10
Bayesian inference (1)
H : Hypothesis
E : Evidence, testimony, or symptom
P(H) : Prior probability with no evidence
P(H|E) : Posterior probability after evidence E has been obtained
Bayesian inference (2)
After E has been obtained, how we should modify the probability of H?
Modification of odds in terms of E
)()()|()|(
EPHPHEPEHP =
)()()|()|(
EPHPHEPEHP =
)()()(
)|()|(
)|()|()|( HO
HPHP
HEPHEP
EHPEHPEHO Eλ===
Prior oddsPosterior odds
2015/10/7
11
Bayesian inference (3)
After another evidence F has been obtained,
)()()(
)|()|(
)|()|(
)|()|()|(
HOHPHP
HFPHFP
HEPHEP
FEHPFEHPFEHO
FEλλ==
∩∩
=∩
Modification factor for each evidence
Hasty doctor example
You received a cancer test. The doctor said that 1 out of 1,000 is the average rate of cancer at your age. The test is accurate and it gives a correct result for 99% of cancer patients and 97% of non cancer patients. Your test result was positive, and the doctor recommended you hospitalization ASAP.
301
999.003.0001.099.0
)()(
)()(
)( ≈××
=++
=+CPCP
CPCP
CO
2015/10/7
12
Three prisoners problem(Monty Hall problem)
Out of three prisoners, two will be executed and one will be freed tomorrow. Prisoner A heard from the jailor that B will be executed. A was delighted that his alive probability has increased from 1/3 to 1/2 with this information, since either A or C will be freed.
31
13/103/12/13/12/13/1
)|()()|()()|()()|()()|(
=⋅+⋅+⋅
⋅=
++=
CbPCPBbPBPAbPAPAbPAPbAP
Component failure example
n Failure statistics of a particular component are collected for every 100 days of operation, and the following data were obtained. Evaluate the failure rate.
Interval 1 2 3 4
Times of failures 6 3 5 6
Average = (6+3+5+6)/4/100 = 0.05 day-1
STD = 0.02 day-1
2015/10/7
13
Probability density function
Cumulative distribution function F(x)n Probability that a random variable X takes
a value not greater than x
Probability density function f(x)
)()(
)()()()(
dxxXxPdxxf
dxxfxFdx
xdFxf x
+≤<≈
== ∫ ∞−
)()( xXPxF ≤<−∞=
Bayesian inference on distribution
: Prior density function of θ: Posterior density function of θ
Bayesian update of the density function after event A has been observed
)(θf)|( Af θ
∫=
θθθθθθ
dfAPfAPAf
)()|()()|()|(
2015/10/7
14
Application of Bayesian approach to component failure example
Poisson distribution
λ : Failure rateT : Interval of observation (100 days)k : Times of failures observed
n Uniform distribution in [0, 0.1] is assumed for f (λ) at the beginning.
Tk
ekTkP λλλ −=!)()|(
Bayesian update of failure rate
0
10
20
30
40
0.00 0.02 0.04 0.06 0.08 0.10
01 (k = 6)2 (k = 3)3 (k = 5)4 (k = 6)
λ (day-1)
P(λ|
A)
2015/10/7
15
Concepts of probability (1)
Normative definition (Pascal)n Fraction of number of cases among the
sample space
Statistical definition (d’Alembert)n Asymptotic value of event frequency
)/(lim nfp nn ∞→=
||/|| UAp =
Concepts of probability (2)
Subjective definition (Bayes)n Degree of individual confidence on
occurrence of the event
Informatics definition (Shannon)n Amount of information that implies
occurrence of the event
pS 2log−=