first, excuse my engish, please. :-)
I am trying to train a Neural Network to be an opponent to a human player in the game HEX (see wikipedia: Hex).
To do this, i have used the same idea as which was already used for TD-Gammon: I am using a free-forward Net with one Input layer (Boardsize*Boardsize+1), one hidden layer (Boardsize*Boardsize/2.3) and one output layer (only 1 row), all sigmoid layers. The net shoud assign the winning-probability to a board-configuration (i.e. 0.0 if it seems, that player 2 will win the game and 1.1 if it seems, that player 1 will win the game).
The board is presented in the following way: 0.0 if the position is taken by player 2, 1.0 if the position is taken by player 1 and 0.5 if the position is empty.
Actually, with a 8x8 Board the net should do the following:
array of 64+1 double values between 0.0 and 1.0 (+1 for the move) -------> value between 0.0 and 1.0
My question is the following: Have you got a clue, how i can determine such values as 'learning rate', the numbers of hidden units, momentum, etc.? What is about two or three hidden layers?
The number of training patterns will be about 1000 and (second net) about 100 000.
Have you got an (heuristic) idea for the values or any other hint?
Thank you for your help,