Can someone help me figure out how to debug my tic-tac-toe attempt?

[rhyolight]

Thanks for explaining your issue to us. I’ll try to help out…

@mgstrein

My plan is to send NUPIC data in pairs of rows, where the first row is the serialization of a game board, and the second indicates the player that won the game.

This already seems wrong to me. Each row of data you send should be a snapshot in time, so the fact that you send the player row after the board state row is going to confuse it. I would suggest you encode the board state and the player in the same row.

Also, what exactly are you trying to predict? Who won? I think a better goal for NuPIC might be to predict the next board state, given previous board states and players info.

@mgstrein

A board can be serialized into a single row of values. These values are a ‘reset’ column, a column for each of the nine positions on a tic-tac-toe board and finally a column that indicates the ‘winner’ of the previous row. A board and its identification can then be expressed with a pair of rows. The first of the two rows has the reset bit set in the first column, the values of the board in the following nine columns, and zero in the winner column. The following row is all zero’s except for the winner column, which indicates the winner - zero, player 1, or player 2.

What if we abandoned the idea of passing in rows of scalar input and thought about a better way to encode the problem space into an SDR?

You could write an encoder that converts each cell on the board (empty, X, or O) to a binary representation. Concatenate them together, and also include a bit for the player. If you train it on hundreds (better thousands) of games, the model should predict the state of the board better than random. I think you’ll have more success this way than trying to shoehorn the semantics of a tic-tac-toe problem space into a bunch of generic scalar encoders.

@mgstrein

My expirment begins by creating a few hundred sample boards. These boards with their results are writtent to a .csv file. I swarm over this csv file to produce a model. Using this model, I feed it more randomly generated boards and their results. I was hoping that after a certain amount of time, given a board my model would always predict the correct value of the winner column in the row that would come next - the result row. However what I am seeing seems to be more or less random.

If you are generating the boards randomly, and generating the moves randomly… then it makes sense to me that you’re getting random predictions back. If there is no pattern in the data, there’s nothing to analyze. You want to look for a catalog or history of tic-tac-toe games played by humans.