XRL-Script

Efficient distributed AutoRL script for any framework

Features

• Distributed training on single machine, multiple CPU/GPUs
• Automatic resource allocation
• TPESampler and Hyperband pruner support
• Automatic testing for stable hyperparameter set

Usage

In rl.py

1. Implement your training logic in train_rl_agent
2. Specify hyperparameter range

In rl-auto-gpu.py

1. Specify minimum and maximum steps per trial, and number of trials
2. Set parallel trials per GPU & parallel envs per trial according to your hardware
3. Set reduction factor (integer). In most situations you can set one so that Hyperband bracket number stays in [4, 6]

Then run rl-auto-gpu.py, the script will find all available gpus and run hyperparameter search in parallel

Dependencies

• Optuna