I have written an AI which plays chess and learns from a dataset of game-states, where each state is a record of a board position and a corresponding result of the game. This causes some inequalities in support for different situations, eg. in my first 1000 games (summing up to around 40000 game-states) there are about 7000 states with all the pieces on board (around 7 states per game). On the other hand only few games of these 1000 had a situation with rook vs 2 bishops ending, summing up to around 10 states out of all 40000.
I am afraid that this inequality might cause the network to overfit to the game openings while underfitting to the endgames.
I am not sure whether this is true and even if this is bad. Openings seem to be more important for positioning and endgames can go deeper into search tree. However I am afraid that due to the inequality my network might prefer to play openings that lead to worse endgames.
My question is:
- is this inequality even a problem for chess?
- if yes - how do I fix it? My idea was to filter the states, but I am not sure how to filter it to keep the support equal