Overview
RLlib is an open-source library built on Ray for reinforcement learning. It provides scalable and fault-tolerant RL workloads with unified APIs for various industry applications, including multi-agent setups, offline data training, and simulator integration. The architecture supports distributed sample collection via EnvRunner actors, loss calculation, and model updating. RLlib integrates with Ray Data for large-scale data ingestion for offline RL and behavior cloning. It supports customization through the RLModule APIs, enabling complex multi-model setups and component sharing between agents. RLlib also provides APIs for compiling and executing accelerated DAGs.