Skip to content
VinF edited this page Jul 12, 2016 · 17 revisions

Full documentation is now available at http://deer.readthedocs.io

Introduction

This package provides a general deep reinforcement Q-learning algorithm. It is based on the original deep Q learning algorithm described in : Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning." Nature 518.7540 (2015): 529-533.

Contrary to the original code, this package provides a more general framework where observations are made up of any number of elements : scalars, vectors and frames (instead of one type of frame only in the above mentionned paper). At each time step, an action is taken by the agent and one observation is gathered (along with one reward). The belief state on which the agent is based to build the Q function is made up of any length history of each element provided in the observation.

Another advantage of this framework is that it is build in such a way that you can easily add up a validation phase that allows to stop the training process before overfitting. This possibility is useful when the environment is dependent on scarce data (e.g. limited time series).

The framework is made in such a way that it is easy to

  • build any environment
  • modify any part of the learning process
  • use your favorite python-based framework to code your own neural network architecture. The provided neural network architectures use Theano (with or without the lasagne library).

Future extensions include:

  • Add planning (e.g. MCTS based when deterministic environment)
  • Several agents interacting in the same environment
  • ...

Please submit any contribution via pull request

This wiki provides the basic concepts to start using the framework. For details, the user can refer to comments in the code itself and to the different working examples. An API documentation will be made available in the future.