NN doesn't want to learn plakoto...

This is starting to get on my nerves. I just recently got into Neural Network programming and I am trying to make a few NNs that will evaluate correctly game boards of backgammon. There are three different game types of backgammon in Greece.

• Fevga
• Plakoto
• Portes

I've started the training with Fevga and it turned out quite good, say above average. And then I moved on to Plakoto...

Man, it does not want to learn how to play this game. I've tried everything. I increased the hidden layers neurons from 40 to 80. I went higher to 120 and 160. I let it train for 10 million games. No good. Always below average on strength of play.

There must be a bug either in coding or even conceptually. I won't give up. I'll get it.