SB3 Trainer Project

Overview

• Showcases how to orchestrate reinforcement-learning experiments with Stable Baselines3 (SB3) inside the AGILab framework.

• Builds a DAG of training tasks (one per agent configuration) and executes them through the distributor so multiple policies can be trained in parallel.

• Highlights how to persist trained artefacts and logs under save_uri so the Experiment and Explore pages can surface them.

Manager (sb3_trainer.sb3_trainer)

• Accepts a list of agent definitions (policy name, hyperparameters) and turns them into a work plan consumed by WorkDispatcher.

• Sets up output directories, handles archive unpacking and keeps the run graph serialised so the Workplan visualiser can display training dependencies.

• Illustrates how to declare non-tabular workloads (function invocations instead of dataframe partitions).

Worker (sb3_trainer_worker.sb3_trainer_worker)

• Executes the generated work plan, instantiates SB3 trainers per agent and persists checkpoints in the configured save_uri.

• Includes pre/post-install scripts to fetch RL-specific assets (Gym wrappers, environments) and test harnesses to validate the training loop.

Assets & Tests

• app_test.py verifies that trainer definitions stream correctly from app_settings.toml into the work plan.

• src/sb3_trainer_worker/run-all-test.py exercises end-to-end training in a controlled environment.

• src/app_settings.toml demonstrates how to describe multiple trainer variants in the AGILab configuration format.

API Reference