mdabros
commented
5 years ago Hi @jameschch,
Reinforcement learning is without doubt an extremely interesting field, with a lot of focus these days. It is also a field I am interested in learning more about, and it could be great to add some methods to SharpLearning. My approach would probably be to start with updating the Neural Net library. The current implementation defenitely needs an overhaul to make it more flexible for new problem types, and also to have it support GPU processing. My current hope is that the TorchSharp project could be used as a backend once it matures a bit more. After that, I would start looking into reinforcement learning algorithms based on neural nets, and continue from there. Since this is a sparetime project, reaching a reinforcement milestone is probably not in the near future though :-).
best regards Mads
jameschch
There is currently support for most of the common (and some less common) ML algorithms in Sharp Learning. However, there does appear to be a lack in the area of Reinforcement Leaning and some might observe these algorithms are beginning to gain some traction.
If there is any appetite for extending into this area, I would propose as an initial baseline provision for QLearning and Sarsa, backed by Epsilon Greedy and Boltzmann approaches. A second stage could then continue with Thompson and UCB1 exploration, and finally the existing Neural Net and ensemble interfaces could probably produce a compound that resembled Deep Q Networks.