These tutorials provide an introductory guide to using AgileRL with PettingZoo. AgileRL’s multi-agent algorithms make use of the PettingZoo parallel API and allow users to train multiple-agents in parallel in both competitive and co-operative environments. This tutorial includes the following:
DQN: Train a DQN agent to play Connect Four through curriculum learning and self-play
MADDPG: Train an MADDPG agent to play multi-agent atari games
MATD3: Train an MATD3 agent to play multi-particle-environment games
AgileRL is a deep reinforcement learning framework focused on streamlining training for reinforcement learning models. Using evolutionary hyper-parameter optimisation (HPO), AgileRL allows users to train models significantly faster and more accurately when compared with traditional HPO techniques. AgileRL’s multi-agent algorithms orchestrate the training of multiple agents at the same time, and benchmarking has shown up to 4x increase in return in a shorter time-frame when compared with implementations of the very same algorithms in other reinforcement learning libraries.