Predictive Game Theory

Conventional noncooperative game theory hypothesizes that the joint (mixed) strategy of a set of reasoning players in a game will necessarily satisfy an “equilibrium concept”. The number of joint strategies satisfying that equilibrium concept has measure zero, and all other joint strategies are considered impossible. Under this hypothesis the only issue is what equilibrium concept is “correct”. This hypothesis violates the first-principles arguments underlying probability theory. Indeed, probability theory renders moot the controversy over what equilibrium concept is correct — while in general there are joint (mixed) strategies with zero probability, in general the set {strategies with non-zero probability} has measure greater than zero. Rather than a firstprinciples derivation of an equilibrium concept, game theory requires a first-principles derivation of a distribution over joint strategies. However say one wishes to predict a single joint strategy from that distribution. Then decision theory tell us to first specify a loss function, a function which concerns how we, the analyst/scientist external to the game, will use that prediction. We then predict that the game will result in the joint strategy that is Bayes-optimal for that loss function and distribution over joint strategies. Different loss functions — different uses of the prediction — give different such optimal predictions. There is no more role for an “equilibrium concept” that is independent of the distribution and choice of loss function. This application of probability theory to games, not just within games, is called Predictive Game Theory (PGT). This paper shows how information theory provides a first-principles argument for how to set a distribution over joint strategies. The connection of this distribution to the bounded rational Quantal Response Equilibrium (QRE) is elaborated. In particular, taking the QRE to be an approximation to the mode of the distribution, correction terms to the QRE are derived. In addition, some Nash equilibria are not approached by any limiting sequence of increasingly rational QRE joint strategies. However it is shown here that every Nash equilibrium is approached with a limiting sequence of joint strategies all of which have non-zero probability. (In general though not all strategies in those sequences are modes of the associated distributions over joint strategies.) ∗D. Wolpert is with NASA Ames Research Center, Moffett Field, CA, 94035 [email protected]