I really want to get into machine learning, someone suggested me in chat to start with Q ML, but I don’t known which challenge could fit with that algorithm. I want to learn Monte Carlo Tree search too, because it seems very reliable on multiple problems/situations. Which multiplayer challenge is better to start ML programming? Where I can try Q tables? Which’s the easiest MCTS-recommended challenge?
1 Like
I don’t think QTables are realistically usable anywhere on here - all the games have too many states for them to work. The simplest thing to learn them is to make a tictactoe implementation (you’d be able to use this for UTTT wood league, but it would be overkill) - it should very quickly learn to become unbeatable. They’re still very useful to learn as it’s the basis for deep Q learning which may be more applicable.
UTTT is by far the best MCTS testing ground - with ok performance (best to use a compiled language) and the most basic version of MCTS (entirely random full rollouts, no game knowledge) you’ll get to the top of gold and then can tweak performance / the mcts algorithm to get into legend.
ML is harder to recommend games for - I had good success on Oware and others have had success on CSB/Bitrunner2048 but I don’t think any of the success has come from simple implementations (eg mine is using alphazero style methods - guiding a MCTS style search with a NN rather than only using the NN). There aren’t any ML frameworks available so you have to implement the inference stage yourself and the code size limitations limit how big your network can be. Resource limitations (single core, no gpu, low time) are also an issue.
Also see https://www.codingame.com/forum/t/neural-network-ressources/1667 for discussion of ML + CSB.
9 Likes
Thank you, very good explanation