search

accord-net / framework

Machine learning, computer vision, statistics and general scientific computing for .NET
http://accord-framework.net
GNU Lesser General Public License v2.1
4.48k stars 1.99k forks source link

Q-Learning with function approximation #276

Open functionsoft opened 8 years ago

functionsoft commented 8 years ago

Hi, Would it be possible to extent the q-learning implementation to use a function approximator instead of a state-action table, and include a sample application?

This would be much more useful for real world reinforcement learning applications.

Thanks,

Mike

cesarsouza commented 8 years ago

Hi there,

Many thanks for the suggestion! Indeed this would be a very welcome addition to the framework. However, the Q-learning implementation currently available in the framework has been inherited from the AForge.NET project and I am not completely familiar with the technique in order to be able to extend it with a function approximator right now. I would need some time to go through the references to understand what would have to be done.

In the meantime, since you seem to have more knowledge on the issue, wouldn't you mind contributing with a minimal implementation that we could develop further afterwards? It doesn't need to be perfect, just enough to get an initial version so I could understand better what would be needed in terms of API, etc. I can also provide a sample application afterwards. If you (or anyone else) would like to tackle on this issue, it would be a huge help!

Best regards, Cesar

lachlann562 commented 7 years ago

You can find a java example here: http://artint.info/demos/rl/sGameFA.html