Random idea that it would be nice to train a neural network that would converge to a game-theory optimal action. Scenario: observed multi-player interaction. Perhaps the interaction is complex enough that a neural network would be useful. Regularie the training process somehow so that the final weights correspond to Nash equilibria?