Google open source new enhanced learning framework Dopamine
Google recently announced open source a new Tensorflow-based framework, Dopamine, designed to provide flexibility, stability and repeatability for novice and experienced RL (Reinforced Learning) researchers. Inspired by reward-motivated behaviour in the brain, the framework reflects a strong historical connection between neuroscience and reinforcement learning research, aimed at achieving speculative research that can drive original discovery.
Dopamine’s feature highlights
Ease of use
- Clearness and simplicity are two key considerations in the design of the framework. The code provided by Google is very compact (about 15 Python files) and well documented, and Google hopes that this simple feature will make it easier for researchers to understand the internal workings of the agent and quickly try new ideas.
Reusability
- Google values the importance of reusability in intensive learning research. To this end, they provide complete test coverage of the code; and follow Machado advice given by scholars, using arcade learning environment (Arcade Learning Environment) standardised empirical evaluation.
Benchmarks
- For new researchers, it is essential to be able to benchmark your ideas quickly based on existing methods. As a result, Google provides complete training data for four agents, including 60 games supported by the Arcade Learning Environment, in the form of Python pickle files (for agents trained with Google Framework) and JSON data files (intelligence for comparison with other frameworks) body). Google also offers an additional website where you can quickly see the agent visualisation training for all 60 games.
