This repository provides a TensorFlow implementation of Self-Imitation Learning (SIL), a reinforcement learning technique that enhances policy learning by allowing agents to learn from their past successful trajectories. It is targeted at researchers and practitioners in reinforcement learning, offering a method to improve sample efficiency and performance on complex tasks.

How It Works

Self-Imitation Learning augments standard policy gradient methods by incorporating a "self-imitation" loss. This loss encourages the agent to imitate its own past high-reward trajectories, effectively learning from its successes. The implementation builds upon the OpenAI Baselines framework, integrating SIL into established algorithms like A2C and PPO.

Quick Start & Requirements

• Primary install/run command: python baselines/a2c/run_atari_sil.py --env FreewayNoFrameskip-v4 (for Atari) or python baselines/ppo2/run_mujoco_sil.py --env Ant-v2 --num-timesteps 10000000 --lr 5e-05 (for MuJoCo).
• Prerequisites: TensorFlow, OpenAI Baselines, Atari environments, MuJoCo environments.
• Links: ICML 2018 Paper

Highlighted Details

• Implements Self-Imitation Learning (SIL) for reinforcement learning.
• Integrates SIL with A2C and PPO algorithms.
• Tested on Atari and MuJoCo benchmark environments.

Maintenance & Community

This project is associated with the ICML 2018 paper "Self-Imitation Learning" by Junhyuk Oh et al. Further community or maintenance activity is not explicitly detailed in the README.

Licensing & Compatibility

The repository's license is not specified in the README. Compatibility with commercial or closed-source projects is therefore undetermined.

Limitations & Caveats

The implementation is based on TensorFlow 1.x, which may require specific environment configurations for compatibility with newer TensorFlow versions or other deep learning frameworks. The README does not detail specific hardware requirements beyond those typical for deep reinforcement learning training.