Introduction
Light-MALib is a Pythonic framework for Population-Based Multi-Agent Reinforcement Learning (PB-MARL). As the name suggests, L-MALib is a simplified version of the MALib codebase which uses Ray as a backend. L-MALib follows the design of an Actor-based RL systems which execute algorithms through message passing between a set of actors deployed on worker node. Similar to MALib, L-MALib possess a high-level agent/evaluator/learner abstraction but drop the burdensome coordinator. This has made the codebase easy to read as well.
Overview of the MALib architecture.
Features
The key features of L-MALib include:
Pytorch-based algorithm implementation: All algorithms implemented in L-MALib are based on PyTorch.
Popular distributed computing RL framework support: L-MALib support multiple distributed computing RL frameworks, including asynchronous
GRF-specific toolkits: L-MALib was initially designed to benchmark Google Research Football multi-agent scenarios. It provides some GRF-specific content that makes GRF-related research easier.