AgentNet is a lightweight Python library for building and training deep reinforcement learning (DRL) and custom recurrent neural networks (RNNs) for Markov Decision Processes. It targets researchers and practitioners looking to prototype DRL models for tasks like game playing, offering flexibility with discrete and continuous control, and supporting various RL algorithms.

How It Works

AgentNet leverages Theano and Lasagne for neural network construction, providing access to standard layers like convolutions, pooling, and dropout. It implements several RL algorithms including Q-learning, SARSA, and Advantage Actor-Critic, with support for N-step learning. The framework is designed for ease of research and prototyping, allowing users to easily swap learning algorithms or integrate custom memory architectures.

Quick Start & Requirements

• Install: [sudo] pip install --upgrade https://github.com/yandexdataschool/AgentNet/archive/master.zip
• Prerequisites: Theano, Lasagne, NumPy, SciPy, OpenAI Gym.
• Docker: justheuristic/agentnet container available.
• Documentation: Tutorials available at classwork.ipynb and documentation pages.

Highlighted Details

• Supports discrete and continuous control problems.
• Implements Q-learning, N-step Q-learning, SARSA, N-step A2C, and N-step DPG.
• Includes boilerplate for custom long-term memory network architectures.
• Examples cover Atari games via OpenAI Gym, synthetic problems, and fact guessing.

Maintenance & Community

The project is under active development, with contributions welcomed. Links to community channels or roadmaps are not explicitly provided in the README.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

AgentNet is described as being under active construction, implying potential for breaking changes. Some examples, like the Atari DQN implementation, are noted as simplistic or suffering from issues like "atari flickering."