Value methods for efficiently solving stochastic games of complete and incomplete information.

Solving multi-agent reinforcement learning problems has proven difficult because of the lack of tractable algorithms. We provide the first approximation algorithm which solves stochastic games with cheap-talk to within ∊ absolute error of the optimal game-theoretic solution, in time polynomial in 1/∊. Our algorithm extends Murray's and Gordon's (2007) modified Bellman equation which determines the set of all possible achievable utilities; this provides us a truly general framework for multi-agent learning. Further, we empirically validate our algorithm and find the computational cost to be orders of magnitude less than what the theory predicts.