VPG

The release contains two small changes:

1. Rename REINFORCE to VPG in order to stay consistent with other libraries. Also, allow VPG to average the gradients over multiple episodes, drastically improving performance in some cases.
2. Tweaked A2C to make it align better with other implementations. In particular, a new n-step buffer was added that is more accurate. There are also some small changes to make sure feature gradients are computed correctly.

A2C

• Add the a2c algorithm
• Support for running parallel environments
• Record the loss function and other information
• Simplify the installation process
• Reenable TravisCI tests

CUDA

1. Add CUDA support
2. Refactor list handling in function approximators
3. Tweak the organization of presets
4. Improve the benchmarks API

DQN

Includes a fully working DQN implementation, as well as a working partial Rainbow. Also includes Actor-Critic, Sarsa, and Reinforce implementations for classic control environments.