A Meta-Learning and Bounded Rationality Framework for Repeated Games in Adversarial Environments

In this work, a meta-learning framework for games between adapting players is proposed. An agent with increased cognitive abilities is augmented with a structure that allows them to identify the way that their opponents learn during the game. This is achieved via approximators that are tuned online leveraging only observed actions from the environment. We show that knowledge of the utilities of the opponents enable asymptotic convergence of the approximation weights. We, then, extend the framework via backpropagation through time such that knowledge of the utilities is not necessary and we show convergence of the errors to a residual set. Finally, simulations of players learning in a penny matching game demonstrate the efficacy of our approach.