Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l -...
Transcript of Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l -...
![Page 1: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/1.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Behavioral Game Theory
Based on joint work with James Wright
Behavioral Game Theory Based on joint work with James Wright, Slide 1
![Page 2: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/2.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 2
![Page 3: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/3.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Computational Problems in Domination
Identifying strategies dominated by a pure strategy
polynomial, straightforward algorithm
Identifying strategies dominated by a mixed strategy
polynomial, somewhat tricky LP
Identifying strategies that survive iterated elimination
repeated calls to the above LP
Asking whether a strategy survives iterated elimination underall elimination orderings
polynomial for strict domination (elimination doesn’t matter)NP-complete otherwise
Behavioral Game Theory Based on joint work with James Wright, Slide 3
![Page 4: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/4.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Rationalizability
Rather than ask what is irrational, ask what is a best responseto some beliefs about the opponent
assumes opponent is rationalassumes opponent knows that you and the others are rational...
Equilibrium strategies are always rationalizable; so are lots ofother strategies (but not everything).
In two-player games, rationalizable ⇔ survives iteratedremoval of strictly dominated strategies.
Behavioral Game Theory Based on joint work with James Wright, Slide 4
![Page 5: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/5.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Formal definition
Definition (Correlated equilibrium)
Given an n-agent game G = (N,A, u), a correlated equilibrium isa tuple (v, π, σ), where v is a tuple of random variablesv = (v1, . . . , vn) with respective domains D = (D1, . . . , Dn), π isa joint distribution over v, σ = (σ1, . . . , σn) is a vector ofmappings σi : Di 7→ Ai, and for each agent i and every mappingσ′i : Di 7→ Ai it is the case that∑
d∈Dπ(d)ui (σ1(d1), . . . , σi(di), . . . , σn(dn))
≥∑d∈D
π(d)ui(σ1(d1), . . . , σ
′i(di), . . . , σn(dn)
).
Behavioral Game Theory Based on joint work with James Wright, Slide 5
![Page 6: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/6.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Existence
Theorem
For every Nash equilibrium σ∗ there exists a correspondingcorrelated equilibrium σ.
This is easy to show:
let Di = Ai
let π(d) =∏
i∈N σ∗i (di)σi maps each di to the corresponding ai.
Thus, correlated equilibria always exist
Behavioral Game Theory Based on joint work with James Wright, Slide 6
![Page 7: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/7.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Remarks
Not every correlated equilibrium is equivalent to a Nashequilibrium
thus, correlated equilibrium is a weaker notion than Nash
Any convex combination of the payoffs achievable undercorrelated equilibria is itself realizable under a correlatedequilibrium
start with the Nash equilibria (each of which is a CE)introduce a second randomizing device that selects which CEthe agents will playregardless of the probabilities, no agent has incentive to deviatethe probabilities can be adjusted to achieve any convexcombination of the equilibrium payoffsthe randomizing devices can be combined
Behavioral Game Theory Based on joint work with James Wright, Slide 7
![Page 8: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/8.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 8
![Page 9: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/9.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Computing CE
∑a∈A|ai∈a
p(a)ui(a) ≥∑
a∈A|ai∈a
p(a)ui(a′i, a−i) ∀i ∈ N, ∀ai, a′i ∈ Ai
p(a) ≥ 0 ∀a ∈ A∑a∈A
p(a) = 1
variables: p(a); constants: ui(a)
we could find the social-welfare maximizing CE by adding anobjective function
maximize:∑a∈A
p(a)∑i∈N
ui(a).
Behavioral Game Theory Based on joint work with James Wright, Slide 9
![Page 10: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/10.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Computing CE
∑a∈A|ai∈a
p(a)ui(a) ≥∑
a∈A|ai∈a
p(a)ui(a′i, a−i) ∀i ∈ N, ∀ai, a′i ∈ Ai
p(a) ≥ 0 ∀a ∈ A∑a∈A
p(a) = 1
variables: p(a); constants: ui(a)
we could find the social-welfare maximizing CE by adding anobjective function
maximize:∑a∈A
p(a)∑i∈N
ui(a).
Behavioral Game Theory Based on joint work with James Wright, Slide 9
![Page 11: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/11.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Why are CE easier to compute than NE?
∑a∈A|ai∈a
p(a)ui(a) ≥∑
a∈A|a′i∈a
p(a)ui(a′i, a−i) ∀i ∈ N, ∀ai, a′i ∈ Ai
p(a) ≥ 0 ∀a ∈ A∑a∈A
p(a) = 1
intuitively, correlated equilibrium has only a single randomizationover outcomes, whereas in NE this is constructed as a product ofindependent probabilities.
To change this program so that it finds NE, the first constraintwould be∑
a∈Aui(a)
∏j∈N
pj(aj) ≥∑a∈A
ui(a′i, a−i)
∏j∈N\{i}
pj(aj) ∀i ∈ N, ∀a′i ∈ Ai.
This is a nonlinear constraint!
Behavioral Game Theory Based on joint work with James Wright, Slide 10
![Page 12: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/12.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 11
![Page 13: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/13.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Fun Game
Guess 95% of the average
Guess 40% of the average
Behavioral Game Theory Based on joint work with James Wright, Slide 12
![Page 14: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/14.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Fun Game
Guess 95% of the average
Guess 40% of the average
Behavioral Game Theory Based on joint work with James Wright, Slide 12
![Page 15: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/15.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Behavioral Game Theory
Behavioral game theory: Aims to extend game theory tomodeling human agents.
There are a wide range of BGT models in the literature.Historically, BGT has been most concerned with explainingbehavior, often on particular games, rather than predicting it.No study compares a wide range of models, considerspredictive performance, or looks at such a large, heterogeneousset of games.
Behavioral Game Theory Based on joint work with James Wright, Slide 13
![Page 16: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/16.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
This Talk
We:
Compared predictive performance of:
Nash equilibrium, plusFour prominent models from behavioral game theoryUsing six experimental datasets from the literature
Bayesian sensitivity analysis:
Yields new insight into existing model (Poisson-CH)Argues for a novel simplification of an existing model(Quantal level-k)
Behavioral Game Theory Based on joint work with James Wright, Slide 14
![Page 17: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/17.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 15
![Page 18: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/18.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Nash equilibrium and human subjects
Nash equilibrium often makes counterintuitive predictions.
In Traveler’s Dilemma: The vast majority of human playerschoose 97–100.
Modifications to a game that don’t change Nash equilibriumpredictions at all can cause large changes in how humansubjects play the game [Goeree & Holt 2001].
In Traveler’s Dilemma: When the penalty is large, people playmuch closer to Nash equilibrium.But the size of the penalty does not affect equilibrium.
Clearly Nash equilibrium is not the whole story.
Behavioral game theory proposes a number of models tobetter explain human behavior.
Behavioral Game Theory Based on joint work with James Wright, Slide 16
![Page 19: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/19.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Nash equilibrium and human subjects
Nash equilibrium often makes counterintuitive predictions.
In Traveler’s Dilemma: The vast majority of human playerschoose 97–100.
Modifications to a game that don’t change Nash equilibriumpredictions at all can cause large changes in how humansubjects play the game [Goeree & Holt 2001].
In Traveler’s Dilemma: When the penalty is large, people playmuch closer to Nash equilibrium.But the size of the penalty does not affect equilibrium.
Clearly Nash equilibrium is not the whole story.
Behavioral game theory proposes a number of models tobetter explain human behavior.
Behavioral Game Theory Based on joint work with James Wright, Slide 16
![Page 20: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/20.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Behavioral game theory models
Themes:1
1 Quantal response: Agents best-respond with high probabilityrather than deterministically best responding.
2 Iterative strategic reasoning: Agents can only perform limitedsteps of strategic “look-ahead”.
One model (QRE) is based on quantal response, two models(Lk, CH) are based on iterative strategic reasoning, and one model(QLk) incorporates both.
1Recall: we restrict attention to unrepeated, simultaneous-move games.Behavioral Game Theory Based on joint work with James Wright, Slide 17
![Page 21: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/21.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
BGT model: Quantal response equilibrium (QRE)
QRE model [McKelvey & Palfrey 1995] parameter: (λ)
Agents quantally best respond to each other.
QBRi(s−i, λ)(ai) =eλui(ai,s−i)∑
a′i∈Aieλui(a
′i,s−i)
Precision parameter λ ∈ [0,∞) indicates how sensitive agentsare to utility differences.
λ = 0 means agents choose actions uniformly at random.As λ→∞, QBR approaches best response.
Behavioral Game Theory Based on joint work with James Wright, Slide 18
![Page 22: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/22.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
BGT models: Iterative strategic reasoning
Level-0 agents choose uniformly at random.
Level-1 agents reason about level-0 agents.
Level-2 agents reason about level-1 agents.
There’s a probability distribution over levels.
Higher-level agents are “smarter”; scarcer
Predicting the distribution of play: weighted sum of thedistributions for each level.
Behavioral Game Theory Based on joint work with James Wright, Slide 19
![Page 23: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/23.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
BGT model: Lk
Lk model [Costa-Gomes et al. 2001] parameters: (α1, α2, ε1, ε2)
Each agent has one of 3 levels: level-0, level-1, or level-2.
Distribution of level [2, 1, 0] agents is [α2, α1, (1− α1 − α2)]
Each level-k agent makes a “mistake” with prob εk, or bestresponds to level-(k − 1) opponent with prob 1− εk.
Level-k agents believe all opponents are level-(k − 1).Level-k agents aren’t aware that level-(k − 1) agents will make“mistakes”.
IBRi,0 = Ai,
IBRi,k = BRi(IBR−i,k−1),
πLki,0 (ai) = |Ai|−1,
πLki,k (ai) =
{(1− εk)/|IBRi,k| if ai ∈ IBRi,k,εk/(|Ai| − |IBRi,k|) otherwise.
Behavioral Game Theory Based on joint work with James Wright, Slide 20
![Page 24: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/24.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
BGT model: Cognitive hierarchy
Cognitive hierarchy model [Camerer et al. 2004] parameter: (τ)
An agent of level m best responds to the truncated, truedistribution of levels from 0 to m− 1.
Poisson-CH: Levels are assumed to have a Poisson distributionwith mean τ .
πPCHi,0 (ai) = |Ai|−1,
πPCHi,m (ai) =
{|TBRi,m|−1 if ai ∈ TBRi,m,0 otherwise.
TBRi,m = BRi
(m−1∑`=0
Pr(Poisson(τ) = `)πPCH−i,`
)
Behavioral Game Theory Based on joint work with James Wright, Slide 21
![Page 25: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/25.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
BGT model: QLk
QLk model [Stahl & Wilson 1994] parameters: (α1, α2, λ1, λ2, λ1(2))
Distribution of level [2, 1, 0] agents is [α2, α1, (1− α1 − α2)]
Each agent quantally responds to next-lower level.
Each QLk agent level has its own precision (λk), and its ownbeliefs about lower-level agents’ precisions (λ`(k)).
πQLki,0 (ai) = |Ai|−1,
πQLki,1 = QBRi(πQLk−i,0 , λ1),
πQLkj,1(2) = QBRj(πQLk−j,0 , λ1(2)),
πQLki,2 = QBRi(πQLk−i,1(2), λ2).
Behavioral Game Theory Based on joint work with James Wright, Slide 22
![Page 26: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/26.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Prediction using Nash equilibrium
We would like to compare BGT models’ predictionperformance to Nash equilibrium.
Unmodified Nash equilibrium is not suitable for predictions:1 Games often have multiple Nash equilibria.2 A Nash equilibrium will often assign probability 0 to some
actions.
We constructed two different Nash-based models to deal withmultiple equilibria:
UNEE (Uniform Nash Equilibrium with Error):
Predict the average of all Nash equilibria, + error.
NNEE (Nondeterministic Nash Equilibrium with Error):
Predict the post-hoc “best” Nash equilibrium, + error.
Both models avoid probability 0 predictions via a tunable errorprobability.
Behavioral Game Theory Based on joint work with James Wright, Slide 23
![Page 27: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/27.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Prediction using Nash equilibrium
We would like to compare BGT models’ predictionperformance to Nash equilibrium.
Unmodified Nash equilibrium is not suitable for predictions:1 Games often have multiple Nash equilibria.2 A Nash equilibrium will often assign probability 0 to some
actions.
We constructed two different Nash-based models to deal withmultiple equilibria:
UNEE (Uniform Nash Equilibrium with Error):
Predict the average of all Nash equilibria, + error.
NNEE (Nondeterministic Nash Equilibrium with Error):
Predict the post-hoc “best” Nash equilibrium, + error.
Both models avoid probability 0 predictions via a tunable errorprobability.
Behavioral Game Theory Based on joint work with James Wright, Slide 23
![Page 28: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/28.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 24
![Page 29: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/29.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Experimental setup: Overview
What do we need to compare predictive models?
1 Metric to measure performance
2 Test to evaluate generalization performance
3 Experimental data describing real human play
Key issue: must set models’ free parameters using data (a trickyoptimization problem), then test generalization performance.
Behavioral Game Theory Based on joint work with James Wright, Slide 25
![Page 30: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/30.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
1. Performance Metric
We score the performance of a model by the likelihood of thetest data:
P(Dtest | M,#»
θ ∗).
To evaluate this score (and make meaningful modelpredictions) we must choose the parameters to maximize thelikelihood of the training data:
#»
θ ∗ = argmax#»θ
P(Dtrain | M,#»
θ ).
Behavioral Game Theory Based on joint work with James Wright, Slide 26
![Page 31: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/31.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
2. Test to evaluate Generalization Performance
We estimate generalization performance using 10-foldcross-validation.
Problem: this estimate may depend upon the particularpartition into folds.
We average over multiple (again, 10) cross-validation runs.
We can then compute 95% confidence interval by assuming at-distribution of these averages [Witten & Frank 2000].
This assumption is validated by the law of large numbers asthe number of cross-validation repetitions grows.
Behavioral Game Theory Based on joint work with James Wright, Slide 27
![Page 32: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/32.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
3. Experimental data
Subjects played 2-player normal form games once each.
Each action by an individual player is a single observation.
Data from six experimental studies, plus a combined dataset:
SW94: 400 observations from [Stahl & Wilson 1994]SW95: 576 observations from [Stahl & Wilson 1995]CGCB98: 1296 observations from [Costa-Gomes et al. 1998]GH01: 500 observations from [Goeree & Holt 2001]CVH03: 2992 observations from [Cooper & Van Huyck 2003]RPC09: 1210 observations from [Rogers et al. 2009]ALL6: All 6974 observations
Behavioral Game Theory Based on joint work with James Wright, Slide 28
![Page 33: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/33.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 29
![Page 34: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/34.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Model comparisons: Nash equilibrium vs. BGT
100
105
1010
1015
1020
1025
1030
1035
ALL6 SW94 SW95 CGCB98 GH01 CVH03 RPC09
Like
lihoo
d im
prov
emen
t ove
r uni
form
dis
tribu
tion
UNEENNEE
Best BGTWorst BGT
UNEE almost always worse than every BGT model(exceptions: GH01, SW95).
Even NNEE worse than QLk and QRE in most datasets.
Behavioral Game Theory Based on joint work with James Wright, Slide 30
![Page 35: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/35.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Model comparisons: Lk and CH vs. QRE
100
105
1010
1015
1020
1025
1030
1035
ALL6 SW94 SW95 CGCB98 GH01 CVH03 RPC09
Like
lihoo
d im
prov
emen
t ove
r uni
form
dis
tribu
tion
LkPoisson-CH
Lk and Poisson-CH performance was roughly similar.No consistent ordering between Lk/Poisson-CH and QRE.
Iterative strategic reasoning and quantal response appear tocapture distinct phenomena.
Behavioral Game Theory Based on joint work with James Wright, Slide 31
![Page 36: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/36.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Model comparisons: Lk and CH vs. QRE
100
105
1010
1015
1020
1025
1030
1035
ALL6 SW94 SW95 CGCB98 GH01 CVH03 RPC09
Like
lihoo
d im
prov
emen
t ove
r uni
form
dis
tribu
tion
LkPoisson-CH
QRE
Lk and Poisson-CH performance was roughly similar.No consistent ordering between Lk/Poisson-CH and QRE.
Iterative strategic reasoning and quantal response appear tocapture distinct phenomena.
Behavioral Game Theory Based on joint work with James Wright, Slide 31
![Page 37: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/37.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Model comparisons: QLk
100
105
1010
1015
1020
1025
1030
1035
ALL6 SW94 SW95 CGCB98 GH01 CVH03 RPC09
Like
lihoo
d im
prov
emen
t ove
r uni
form
dis
tribu
tion
QLkBest non-QLk
Worst non-QLk
We would expect a model with both iterative and quantalresponse components to perform best.That is the case: QLk is the best predictive model on almostevery dataset.
Behavioral Game Theory Based on joint work with James Wright, Slide 32
![Page 38: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/38.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Talk Overview
1 Recap
2 Computing Correlated Equilibria
3 Behavioral Game Theory
4 Models of Human Behavior in Simultaneous-Move Games
5 Experimental Setup
6 Comparing our Models in Terms of Predictive Performance
7 Digging Deeper: Bayesian Analysis of Model Parameters
Behavioral Game Theory Based on joint work with James Wright, Slide 33
![Page 39: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/39.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Taking Stock of What We Have Done
Take-home message so far:
QLk is the best of the models for prediction.
For the rest of the talk, we will concentrate on gaining adeeper understanding of QLk. We will:
use Bayesian methods to understand QLk’s parameter space
determining which ranges of values parameters can takeidentifying the most important parameters
indulge ourselves by digressing to consider Poisson-CH ©
Behavioral Game Theory Based on joint work with James Wright, Slide 34
![Page 40: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/40.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Refresher: QLk’s Parameters
QLk has 5 different parameters:
α1: Proportion of level-1 agents.
α2: Proportion of level-2 agents.
λ1: Precision of level-1 agents.
λ2: Precision of level-2 agents.
λ1(2): Level-2 agents’ belief about level-1 agents’ precision.
πQLki,0 (ai) = |Ai|−1,
πQLki,1 = QBRi(πQLk−i,0 , λ1),
πQLkj,1(2) = QBRj(πQLk−j,0 , λ1(2)),
πQLki,2 = QBRi(πQLk−i,1(2), λ2).
Behavioral Game Theory Based on joint work with James Wright, Slide 35
![Page 41: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/41.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Bayesian Sensitivity Analysis
Two questions:1 How sure are we about the parameter values we fit?
That is, how strongly does the data argue for particularparameter values?
2 How important are the different parameters?
We say that a parameter is important if the model’spredictions are substantially degraded if we change its value.
Behavioral Game Theory Based on joint work with James Wright, Slide 36
![Page 42: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/42.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Bayesian Sensitivity Analysis
Two questions:1 How sure are we about the parameter values we fit?
That is, how strongly does the data argue for particularparameter values?
2 How important are the different parameters?
We say that a parameter is important if the model’spredictions are substantially degraded if we change its value.
Behavioral Game Theory Based on joint work with James Wright, Slide 36
![Page 43: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/43.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions
Maximum likelihood only tells us the most likely parametersetting, given the data.
The posterior distribution over parameter settings describesthe relative probability (normalized likelihoods) of all possibleparameter settings.
Individual parameters can be analyzed by inspecting themarginal posterior distribution.
Flat distributions indicate uncertainty about parameter values.Sharp distributions indicate a high degree of certainty.
Behavioral Game Theory Based on joint work with James Wright, Slide 37
![Page 44: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/44.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions: Monte Carlo sampling
The posterior distribution rarely has an analyticrepresentation.
We use Monte Carlo sampling to draw an approximate sampleof values from the joint posterior distribution.
Expectations taken over these approximate samples areunbiased estimators of the true expectations.
Behavioral Game Theory Based on joint work with James Wright, Slide 38
![Page 45: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/45.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Warm-up: Poisson-CH
Regarding the single parameter (τ) for the Poisson-CH model:
“Indeed, values of τ between 1 and 2 explain empiricalresults for nearly 100 games, suggesting that a τ value of1.5 could give reliable predictions for many other gamesas well.” [Camerer et al. 2004]
Behavioral Game Theory Based on joint work with James Wright, Slide 39
![Page 46: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/46.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Warm-up: Poisson-CH’s Posterior Distribution
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.7 0.72 0.74 0.76 0.78 0.8 0.82 0.84
Cum
ulat
ive
prob
abilit
y
τ
Our analysis gives 99% posterior probability that the best value ofτ is 0.8 or less.
Behavioral Game Theory Based on joint work with James Wright, Slide 40
![Page 47: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/47.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions: QLk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ1λ2λ1(2)
Some surprises:1 α1, α2: Best fits predict more level-2 agents than level-1.2 λ1 is not very identifiable from data; multimodal.3 λ1, λ2: Level-2 agents have lower precision than level-1 agents.4 λ1, λ1(2): Level-2 agents’ beliefs are very wrong.
Behavioral Game Theory Based on joint work with James Wright, Slide 41
![Page 48: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/48.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions: QLk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ1λ2λ1(2)
Some surprises:1 α1, α2: Best fits predict more level-2 agents than level-1.
2 λ1 is not very identifiable from data; multimodal.3 λ1, λ2: Level-2 agents have lower precision than level-1 agents.4 λ1, λ1(2): Level-2 agents’ beliefs are very wrong.
Behavioral Game Theory Based on joint work with James Wright, Slide 41
![Page 49: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/49.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions: QLk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ1λ2λ1(2)
Some surprises:1 α1, α2: Best fits predict more level-2 agents than level-1.2 λ1 is not very identifiable from data; multimodal.
3 λ1, λ2: Level-2 agents have lower precision than level-1 agents.4 λ1, λ1(2): Level-2 agents’ beliefs are very wrong.
Behavioral Game Theory Based on joint work with James Wright, Slide 41
![Page 50: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/50.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Posterior distributions: QLk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ1λ2λ1(2)
Some surprises:1 α1, α2: Best fits predict more level-2 agents than level-1.2 λ1 is not very identifiable from data; multimodal.3 λ1, λ2: Level-2 agents have lower precision than level-1 agents.4 λ1, λ1(2): Level-2 agents’ beliefs are very wrong.
Behavioral Game Theory Based on joint work with James Wright, Slide 41
![Page 51: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/51.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Sensitivity measures: Main and total effect
Definition
The main effect of a parameter θj is the percentage of variance inthe prediction that is reduced when θj is fixed to its true value.
Problem: If a parameter has most of its influence frominteractions with other parameters, it will have a low main effect.
Could compute second-order effects, third-order, . . .
There are exponentially many of these!
Definition
The total effect of a parameter θj is the sum of all main andhigher-order effects that θj participates in.
We estimate both of these quantities using Monte Carlo sampling.
Behavioral Game Theory Based on joint work with James Wright, Slide 42
![Page 52: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/52.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Parameter importance: QLk
0
0.2
0.4
0.6
0.8
1
α→ λ1 λ2 λ1(2)
Porti
on o
f var
ianc
e
Parameter
Main effectTotal effect
High interaction effects for all parameters.
Precision parameters influence mostly through interactions.
Proportion parameters ( #»α) about twice as important asprecision parameters (λ1, λ2, λ1(2)).
Behavioral Game Theory Based on joint work with James Wright, Slide 43
![Page 53: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/53.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Homogeneous QLk
QLk is not very sensitive to its individual precision parameters.
The precision parameters are also hard to identify.
Would a single precision parameter λ serve just as well?
Definition (Homogeneous QLk model)
πHQLki,0 (ai) = |Ai|−1,
πHQLki,1 = QBRi(πHQLk−i,0 , λ),
πHQLki,2 = QBRi(πHQLk−i,1 , λ).
Behavioral Game Theory Based on joint work with James Wright, Slide 44
![Page 54: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/54.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Homogeneous QLk: Parameter importance
0
0.2
0.4
0.6
0.8
1
α→ λ
Porti
on o
f var
ianc
e
Parameter
Main effectTotal effect
Virtually no interaction effects.
Proportion parameters ( #»α) still about twice as important asprecision (λ).
Behavioral Game Theory Based on joint work with James Wright, Slide 45
![Page 55: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/55.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Thinking back to QLk
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ1λ2λ1(2)
Recall...
α1, α2: Best fits predict more level-2 agents than level-1.
λ1 is not very identifiable from data; multimodal.
λ1, λ2: Level-2 agents have lower precision than level-1 agents.
λ1, λ1(2): Level-2 agents’ beliefs are very wrong.
Behavioral Game Theory Based on joint work with James Wright, Slide 46
![Page 56: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/56.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Homogeneous QLk: Posterior distribution
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
Cum
ulat
ive
prob
abilit
y
Level proportions
α1α2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
0 1 2 3 4 5 6
Cum
ulat
ive
prob
abilit
y
Precisions
λ
More agents of type 1 than 2.
More confident identification of precision (λ):
Unimodal;Narrower confidence region.
Behavioral Game Theory Based on joint work with James Wright, Slide 47
![Page 57: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/57.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Homogeneous QLk: Performance
100
105
1010
1015
1020
1025
1030
1035
ALL6
Like
lihoo
d im
prov
emen
t ove
r uni
form
dis
tribu
tion
UNEENNEE
LkPoisson-CH
QREQLk
H-QLk
Performance of H-QLk and QLk are similar; H-QLk stilloutperforms all other models.
Behavioral Game Theory Based on joint work with James Wright, Slide 48
![Page 58: Behavioral Game Theory - University of British Columbiakevinlb/teaching/cs532l - 2011-12/lectures/lect6b.pdf · RecapComputing CEBehavioral Game TheoryModelsExperimental SetupModel](https://reader034.fdocuments.us/reader034/viewer/2022042218/5ec36702fe64a93a3e4f92fb/html5/thumbnails/58.jpg)
Recap Computing CE Behavioral Game Theory Models Experimental Setup Model Comparisons Bayesian Analysis
Summary
Compared predictive performance of four BGT models.
BGT models typically predict human behavior better thanNash equilibrium-based models.QLk has best performance of the four.
Bayesian sensitivity analysis of parameters.
Best parameter for Poisson-CH is likely much lower thanpreviously thought.Parameters for QLk are counterintuitive, hard to identify,interaction-laden.Using a single precision for all agents yields better moreintuitive parameter values, similar predictive performance.
Behavioral Game Theory Based on joint work with James Wright, Slide 49