tensorforce/tensorforce

What it solves

Tensorforce is a deep reinforcement learning (RL) framework designed to make the application of RL in research and practice more straightforward and flexible. It addresses the complexity of implementing RL algorithms by providing a modular, component-based design that separates the RL logic from the specific application environment.

How it works

Built on top of TensorFlow, the entire RL logic—including control flow—is implemented as TensorFlow models. This allows for portable computation graphs and easier deployment. The framework uses a modular approach where users can combine different network layers, memory types, policy distributions, and optimization algorithms to build a variety of popular RL models (such as DQN, PPO, and A3C).

Who it’s for

Researchers and practitioners who need a flexible, TensorFlow-based library for implementing and implementing deep reinforcement learning agents across various environments.

Highlights

• Modular Design: Components like network layers, memory types, and reward estimation are highly configurable and generally applicable.
• Crossover Environment Support: Includes adapters for OpenAI Gym, CARLA, ViZDoom, Arcade Learning Environment, and others.
• Full TensorFlow Integration: Logic is implemented in TensorFlow to facilitate portable computation graphs.
• Broad Algorithm Support: Capable of replicating popular models like Q-learning, Policy Gradient, and Trust Region Policy Optimization.
• Parallel Execution: Supports parallelized execution of multiple environments using Python's multiprocessing and socket.