ゲームにおける学習プレイヤーのダイナミクス

抄録

Behavior of cognitive game players who build the opponent's internal model is studied. Internal models are constructed by the recurrent neural network, and the iterated prisoner's dilemma game is performed. The complicated transients of actions are observed before the stable mutually defecting equilibrium is reached. These chaotic dynamics reflect the dynamical and high-dimensional rugged landscape of the internal model space. A possible world analysis reveals the other deep problem in the game thoery, i.e. uncertainty of games. Differences in a payoff matrix show that different matrix gives different possible worlds in behind. Some possible worlds can sustain a mutually cooperative state with mutually believing that the other player is playing Tit for Tat.