Using fast.ai library to learn Connect 4

I want to try to use the fastai library to learn to play the game of connect4.
I do not believe this has been done before successfully although the following
http://www.gm.fh-koeln.de/~konen/Publikationen/ppsn2012_RL-CFour.pdf
describes another approach that was successful.

The game has been solved by brute force so I have been able to get all the board positions for the first 10 moves and the whether they are win, lose or draw.

As a first step I have used this information to train a fully connected NN which gets about 86% accuracy.
The next step is to try using a convolutional network.

So my questions are:

  • is it likely that using Resnet50 trained on the CIFAR-10 dataset is going to be useful at all?
  • the ‘images’ of the board are 7x6 pixels so extremely small is this a problem?
  • the ‘images’ are grayscale, ie only one channel. How should I handle that?
  • finally in addition to the ‘image’ of the board there is one extra piece of information which is who’s move it is.
    I think this should be provided after the convolutional layers to the final fully connected layers. Is there a way to do this using the fast.ai library?

Alan