Tested with: Python 2.7 & 3.7
Runs the Minimax algorithm and trains an ANN with the output.
Uses Dense Layers, Convolutional-Layers would probably (99.9995%) be better.
You need only to adjust the LEN_ITER & run. Played boards may be printed and model be saved.
After some iterations (>100k-1000k i guess) it hopefully converges to a quite good solution.. Or not, i didn't try it that far, just for ~50k, and it did quite good moves, but still some(probably many) bad moves..
The ANN-Topology maybe needs an adjustment as well, it's always trial and error isn't it :)