I'm using Python to write a simple board game (based off generals.io) and have the system learn strategy from observing how the user plays; How can this be formulated as a machine-learning problem?
Description
The game is won when one player captures the other's general. Usually, the winning player has a much larger army than the loser so I thought army size could be used to weight the outcome, but it is also possible to win with a much smaller army by using some more complex strategies, so essentially the outcome is a binary win-lose.
Some types of data available to the program are a turn count, the size of the opponents army and a 2-dimensional list, where each element is a dictionary representing a square of the board. Each dictionary holds who owns the square, what type of square it is (starting position, obstacle etc) and what value the square holds (the size of the army on it).
Current approach
I've looked around an a MCTS seems like it would be a good option, but this is not a turn based game, and the for the most part the other player's positions and moves are unknown (until a conflict occurs).