This repository provides a Soft Actor-Critic (SAC) implementation for training maximum entropy policies in continuous domains, targeting researchers and practitioners in deep reinforcement learning. It offers a TensorFlow-based framework for maximum entropy RL, enabling more robust and diverse policy learning.

How It Works

The framework implements the Soft Actor-Critic algorithm, which optimizes a policy to maximize both expected return and policy entropy. This encourages exploration and leads to more robust policies. The implementation leverages TensorFlow for its deep learning components and is designed for continuous control tasks.

Quick Start & Requirements

• Installation: Via Docker (docker-compose up) or local setup requiring cloning rllab and adding it to PYTHONPATH.
• Prerequisites: Docker, Docker Compose, and a Mujoco license (mjkey.txt). Local setup requires specific Mujoco binaries (mjpro131_linux.zip or mjpro131_osx.zip) and conda for environment management.
• Setup: Docker setup is straightforward if dependencies are met. Local setup involves manual file copying and environment variable configuration.
• Links: rllab, rlkit (PyTorch)

Highlighted Details

• Implements the SAC algorithm from the ICML 2018 paper.
• Supports multiple continuous control environments (ant, walker, swimmer, half-cheetah, humanoid, hopper).
• Includes scripts for training agents and simulating policies.

Maintenance & Community

This repository is no longer maintained. Users are directed to the new softlearning package. Key contributors include Tuomas Haarnoja, Vitchyr Pong, and Kristian Hartikainen. The work was supported by Berkeley Deep Drive.

Licensing & Compatibility

The repository's license is not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

This repository is explicitly marked as no longer maintained. The simulation script is noted to fail with the Docker installation due to missing display dependencies. Users should migrate to the softlearning package for current support.