This repository provides reinforcement learning agents implemented in Keras for OpenAI Gym environments, targeting researchers and practitioners in deep reinforcement learning. It aims to simplify experimentation with different learning configurations and neural network architectures for value function approximation.

How It Works

KeRLym offers Policy Gradient (PG) and Deep Q-Network (DQN) agents. The DQN agent utilizes Keras neural networks for Q-function approximation, supporting concurrent actor-learners. It allows for flexible network customization, including custom Keras models, and incorporates features like epsilon-greedy exploration with decay and experience replay.

Quick Start & Requirements

• Install: pip install gym and pip install gym[atari]. For Atari, ensure cmake is installed (apt-get install cmake).
• Dependencies: Keras, Theano/TensorFlow, OpenAI Gym.
• Usage: Run agents via command line (./run_pong.sh) or programmatically using the kerlym Python API.
• Documentation: https://arxiv.org/abs/1605.09221

Highlighted Details

• Supports Policy Gradient and DQN algorithms.
• Allows custom Keras network architectures (e.g., CNNs, LSTMs).
• Includes configurable parameters for exploration, discount factor, learning rate, and network specifics.
• Offers plotting of learning statistics and optional game rendering.

Maintenance & Community

This repository is explicitly marked as DEPRECATED. The authors recommend using more actively maintained projects like RLLAB.

Licensing & Compatibility

The README does not explicitly state a license. However, the project's dependencies (Keras, TensorFlow, Theano) have their own licenses. Compatibility for commercial use or closed-source linking would require clarification of the project's specific license.

Limitations & Caveats

The project is deprecated and may not be actively maintained or debugged. Users are advised to use alternative, more up-to-date libraries.