Strategy Learning in 3x3 Games by Neural Networks

This paper presents a neural network based methodology for examining the learning of game-playing rules in never-before seen games. A network is trained to pick Nash equilibria in a set of games and then released to play a larger set of new games. While faultlessly selecting Nash equilibria in never-before seen games is too complex a task for the network, Nash equilibria are chosen approximately 60% of the times. Furthermore, despite training the network to select Nash equilibria, what emerges are endogenously obtained bounded-rational rules which are closer to payoff dominance, and the best response to payoff dominance.