So I have 4GB of turn-by-turn data for many games of a particular strategy game. It appears that most people interested in using ML to build an AI for turn-based games use reinforcement learning to build a model on the fly.

Since I already have really good data, can I use supervised learning to solve this task?

EDIT: I was considering using regression to assign a score to a given action based on its likelihood of eventually resulting in a win; is this the right way to think about it?


1 Answer 1


Maybe the correct way of addressing this is by making sub optimizations of every step, even though it could be done by regression, I would suggest decision trees.

You have and advantage: A game is made of discrete steps, so in every moment you can "stop" and decide the best move based on your (possibly comprehensive) history of moves.

Supervised learning vs reinforcement learning for a simple self driving rc car

  • $\begingroup$ Could both regression and decision trees be used? I feel like the output of the prediction of a given turn needs to be a continuous value. $\endgroup$ Apr 11, 2019 at 18:29
  • $\begingroup$ Yes, you could use both of them. I don't know which game your AI is playing but is difficult to think in a game which has a continuos output of a move. Think in Chess (for example), the queen moves n blocks and every block has its own output. Regression is not enough to get the no-linear scoring schemas of most turn-based games. $\endgroup$ Apr 11, 2019 at 20:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.