[Paper Review] A Predictive Theory of Games

David H. Wolpert|ArXiv.org|Dec 8, 2005

Game Theory and Applications71 references3 citations

TL;DR

This paper introduces Predictive Game Theory (PGT), a first-principles framework that replaces equilibrium concepts with a Bayesian-probabilistic approach to predicting joint strategies in games. By using information-theoretic priors and decision theory, PGT derives a distribution over joint strategies, showing that Nash equilibria are not necessarily predictive and that quantal response equilibria emerge as approximations to the mode of this distribution, with correction terms derived via entropy maximization.

ABSTRACT

Conventional noncooperative game theory hypothesizes that the joint strategy of a set of players in a game must satisfy an "equilibrium concept". All other joint strategies are considered impossible; the only issue is what equilibrium concept is "correct". This hypothesis violates the desiderata underlying probability theory. Indeed, probability theory renders moot the problem of what equilibrium concept is correct - every joint strategy can arise with non-zero probability. Rather than a first-principles derivation of an equilibrium concept, game theory requires a first-principles derivation of a distribution over joint (mixed) strategies. This paper shows how information theory can provide such a distribution over joint strategies. If a scientist external to the game wants to distill such a distribution to a point prediction, that prediction should be set by decision theory, using their (!) loss function. So the predicted joint strategy - the "equilibrium concept" - varies with the external scientist's loss function. It is shown here that in many games, having a probability distribution with support restricted to Nash equilibria - as stipulated by conventional game theory - is impossible. It is also show how to: i) Derive an information-theoretic quantification of a player's degree of rationality; ii) Derive bounded rationality as a cost of computation; iii) Elaborate the close formal relationship between game theory and statistical physics; iv) Use this relationship to extend game theory to allow stochastically varying numbers of players.

Motivation & Objective

• To challenge the conventional game-theoretic reliance on equilibrium concepts as the sole valid outcomes.
• To address the inconsistency of equilibrium-based models with foundational principles of probability theory.
• To develop a first-principles method for deriving a distribution over joint strategies using information theory.
• To show that Nash equilibria are not always approachable through rationality limits, and that restricting support to them is often impossible.
• To provide a model-independent quantification of rationality in player behavior using statistical mechanics analogies.

Proposed method

• Uses Bayesian inference to derive a posterior distribution over joint mixed strategies given game structure and prior information.
• Applies maximum entropy (MaxEnt) principle with constraints on expected payoff to derive a distribution over strategies.
• Models player behavior via the Boltzmann distribution, linking rationality to inverse temperature β in a statistical mechanics framework.
• Derives correction terms to the Quantal Response Equilibrium (QRE) by expanding the entropy-maximizing distribution around its mode.
• Uses decision theory to select Bayes-optimal predictions based on loss functions, eliminating the need for equilibrium concepts.
• Extends game theory to stochastic player counts by modeling fluctuations in population size, analogous to statistical physics ensembles.

Experimental results

Research questions

• RQ1 Why is the conventional reliance on equilibrium concepts inconsistent with probability theory’s foundational axioms?
• RQ2 How can a distribution over joint strategies be derived from first principles in game theory?
• RQ3 What is the formal relationship between the Quantal Response Equilibrium and the mode of the derived distribution over strategies?
• RQ4 Can every Nash equilibrium be approached as a limit of increasingly rational QRE strategies, and if not, why?
• RQ5 How can rationality in player behavior be quantified independently of specific models, using information-theoretic measures?

Key findings

• The set of joint strategies with non-zero probability under the derived distribution has measure greater than zero, contradicting the zero-measure assumption of equilibrium concepts.
• Nash equilibria are not always approachable via sequences of increasingly rational QRE strategies, and in many games, restricting the support of the distribution to Nash equilibria is mathematically impossible.
• The Quantal Response Equilibrium (QRE) is shown to be an approximation to the mode of the information-theoretic distribution over joint strategies, with correction terms derived from higher-order expansions.
• Every Nash equilibrium can be approached as a limit of joint strategies that all have non-zero probability, though not necessarily as modes of the associated distributions.
• The framework provides a model-independent quantification of rationality through the inverse temperature parameter β, which emerges naturally from entropy maximization under expected payoff constraints.
• The extension to stochastic player numbers leads to corrections of the replicator dynamics in evolutionary game theory, framed as fluctuations in population size within a statistical mechanics analogy.