Method of Engineering Communication Network Architecture for Improved Distributed Deep Reinforcement Learning

In some implementations of this invention, the performance of a network of reinforcement learning agents is maximized by optimizing the communication topology between the agents for the communication of gradients, weights or rewards. For instance, a sparse Erdos-Renyi network may be employed, and network density may be selected in such a way as to maximize reachability and to minimize homogeneity. In some cases, a sparse network topology is employed for massively distributed learning, such as across entire fleets of autonomous vehicles or mobile phones that learn from each other instead of requiring a master to coordinate learning.