Decentralized Multi-Agent Advantage Actor-Critic
preprintposted on 2022-02-22, 00:56 authored by Scott BarnesScott Barnes
We present a decentralized advantage actor-critic algorithm that utilizes learning agents in parallel environments with synchronous gradient descent. This approach decorrelates agents’ experiences, stabilizing observations and eliminating the need for a replay buffer, requires no knowledge of the other agents’ internal state during training or execution, and runs on a single multi-core CPU.
Email Address of Submitting Authorscottgbarnes@gwu.edu
Submitting Author's InstitutionThe George Washington University
Submitting Author's Country
- United States of America