This is a simple JavaScript implementation of a reinforcement algorithm which simply weights states 'visited' in order to win, and reduces weights visited on a loss. Draw has no effect.

The permutations of options are in the billions for a full board, so the board can be set to smaller (4x4 is the smallest possible).

Not very.

The permutations are still pretty large, so it hasn't seen the same conditions sufficiently to make a judgement to lower or increase weights.

This is not very fast, and doesn't have any hashing of the board to consider similar states such as a mirror image or shifted left/right.

You can view this in this repository GitHub Pages.

The JS will run 2000 iterations and then show the moves the players are making for each run from there.

There is an second model added, which uses the surrounding neighbors as a view of state which will reduce the states and attempts to do some simplistic model which will allow same shapes to match in different positions.

The third model is a rough TensorFlow.js implementation. This doesn't do reinforcement, and just simply provides the grid input. This is trained with 500 totally random plays, and the training data is just the last state of the board with the winning move.

There is some bug somewhere which allows a play a non-existent column on the far right (or possible a negative one on the left) which crops up, but I can't find what it is. That's what I get for not writing any tests for this.

Why not?

This was inspired by the 2019 Royal Institution Christmas Lectures which had a match-box powered Noughts and Crosses player which used the same algorithm (but used coloured beads for moves, rather than an array of columns as I have).