This paper considers a distributed reinforcement learning problem in the presence of Byzantine agents. The system consists of a central coordinating authority called "master agent" and multiple computational entities called "worker agents". The master agent is assumed to be reliable, while, a small fraction of the workers can be Byzantine (malicious) adversaries. The workers are interested in cooperatively maximize a convex combination of the honest (non-malicious) worker agents' long-term returns through communication between the master agent and worker agents. A distributed actor-critic algorithm is studied which makes use of entry-wise trimmed mean. The algorithm's communication-efficiency is improved by allowing the worker agents to send only a scalar-valued variable to the master agent, instead of the entire parameter vector, at each iteration. The improved algorithm involves computing a trimmed mean over only the received scalar-valued variable. It is shown that both algorithms converge almost surely.
展开▼