Repeated Games, Automata, and Growing Strategy Space

• Autores: Abraham Neyman, Daijiro Okada
• Localización: Abstracts of the Fifth Spanish Meeting on Game Theory and Applications / coord. por Jesús Mario Bilbao Arrese , Francisco Ramón Fernández García , 2002, ISBN 84-472-0733-1, pág. 72
• Idioma: inglés
• Texto completo no disponible (Saber más ...)
• Resumen

  • A (pure) strategy in a repeated game is a mapping from histories, or, more generally, signals, to actions. We view the implementation of such a strategy as a computational procedure and attempt to capture in a formal model the following intuition: as the game proceeds, the amount of information (history) to be taken into account becomes large and the �computational burden� becomes increasingly heavy.

    The number of strategies in repeated games grows double-exponentially with the number of repetitions. This is due to the fact that the number of histories grows exponentially with the number of repetitions and also that we count strategies that map histories into actions in all possible ways. Any model that captures the intuition mentioned above would impose some restriction on the way the set of strategies available at each stage expands.

    We point out that existing measures of complexity of a strategy, such as the number of states of an automaton that represents the strategy needs to be refined in order to capture the notion of growing strategy space. Thus we propose a general model of repeated game strategies which are implementable by automata with growing number of states with restrictions on the rate of growth. With such model, we revisit some of the past results concerning the repeated games with finite automata whose number of states are bounded by a constant, e.g., Ben-Porath (1993) in the case of two-person infinitely repeated games.

    In addition, we study an undiscounted infinitely repeated two-person zerosum game in which the strategy set of player 1, the maximizer, expands �slowly� while there is no restriction on player 2�s strategy space. Our main result is that, if the number of strategies available to player 1 at stage n grows subexponentially with n, then player 2 has a pure optimal strategy and the value of the game is the maxmin value of the stage game, the lowest payoff that player 1 can guarantee in one-shot game. This result is independent of whether strategies can be implemented by automaton or not. This is a strong result in that an optimal strategy in an infinitely repeated game has, by definition, a property that, for every c, it holds player 1�s payoff to at most the value plus c after some stage.