This module provides boilerplate code for training RL systems in parallel with ray. ParRL is a PyTorch focused modular library for RL projects.
The system architecture depends on two entities, the Learner and the Gatherer.
-
Learners Learners are responsible for agent training and for maintaining Gatherers. The Learner interacts with environment experience through the use of a ReplayBuffer. Typically, a Learner will host multiple Gatherers.
-
Gatherers Gatherers are responsible for gathering experiences that can be added to the Learner's ReplayBuffer. A Gatherer is a Ray Actor that houses a copy of the Learner's agent and an Environment.