Deep Reinforcement Learning

What is Deep RL?


Deep RL is currently in heavy research has recently led to some exciting successes in AI. Deep RL uses the same deep learning algorithms that have led to huge gains in AI in the last decade plus and applies them to the field of Reinforcement Learning.

There is a huge amount of resources online now on Deep RL. The Sutton and Barto Book is excellent place to start to learn core Reinforcement Learning concepts and the book is freely available online.

Total AI's Planned Deep RL System


Unity has an amazing Open Source ml-agents project which implements state-of-the-art Deep RL. TAI's Deep RL will tie into this system.

The Planner Type will be similar to Utility AI is that it will just return all available Mappings with no chaining. There will be one discrete branch of possible Actions. The Deep RL Decider Type will then mask the Actions based on the ones that are available. It will be responsible for sending the possible Actions and Observations to ml-agents. The rewards will be customizable but will default to the amount of drive utility reduced.

The eventual long term goal for Deep RL is to have Total AI's system automatically converted into a similar network to the Network OpenAI created for their Dota2 Agents: OpenAI Dota2 Network Model

If you are interesting in contributing to the planning and development of Deep RL for TAI please see the Contribute Documentation.