Open ThorvaldAagaard opened 11 months ago
Wouldn't it make sense to create two NNs for declarer play - one for NT one for suit contracts? The strategy and also the rules are quite different. In the code we would need to implement which NN is chosen depending on the final contract.
Yes, I have just implemented that, it will be committed when I am done testing. It doesn't seem to make a big difference. As usual finding good training data is difficult.
Long time ago I made the same split for opening leads against suit and NT, and that was clearly better
This is one of the examples of boards, where it looks like BENs neural network for play is flawed, as HK scores 0.374 - better than a small heart.
Later on this board there is another very strange situation
The double dummy gives the same result for C3 and CT, and the neural network then suggest C3 as the best play (but not by much).