3 ms·
It would be interesting to see these classical methods compared to a simple RL method that optimizes a tiny neural net or decision tree which is given access to
by markisus 3y ago
It would be interesting to see these classical methods compared to a simple RL method that optimizes a tiny neural net or decision tree which is given access to the map, player, and monster positions. I think it would cost a tiny bit more compute but have fewer annoying edge cases.
- Calavar 3y agoI suspect that it would have a similar number of edge cases, but they would be harder to interpret and reproduce. That's just been my experience with moving from unlearned algorithms to deep learning in general.
- qumpis 3y agoI haven't seen RL with decision trees! it sounds really interesting. Any classic results worth looking into?
- markisus 3y agoThis is a relatively recent paper you might find interesting. https://arxiv.org/abs/2204.03771 https://arxiv.org/abs/2204.03771 A random forest is also a universal function approximator so anything a neural net can do, so can a random forest (in theory). In practice, neural nets are easier for modern hardware to optimize while I think trees incur computational overhead due to branchiness.
- malux85 3y agoUsually what happens in cases like this is that the RL network ends up learning a slightly crappier version of A* Or finds a bug in the environment and learns to teleport, those are fun too