machine learning - Ideal Input for Neural Network For The Game Checkers -

i'm new in neural networks , i'm designing feed forward neural network learn play game checkers. input, board has given , output should give chance win , lose. how can ideal transformation of checkers board row of numbers input? there 32 possible squares , 5 different possibility (king or piece of white or black player , free position) on each square. if provide input unit each possible value each square, 32 * 5. option that:

  free position: 0 0    piece of white: 0 0.5 && king piece of white: 0 1    piece of black: 0.5 1 && king piece of black: 1 0 

in case, input length 64 i'm not sure 1 give better result?

i've done sort of thing tic-tac-toe. there several ways represent this. 1 of common ttt have input , output represent entire size of board. in ttt becomes 9 x hidden x 9. input of -1 x, 0 none, 1 o. input neural network current state of board. output desired move. whatever output neuron has highest activation going move.

propagation training not work here because not have finite training set. simulated annealing, pso, or score function ideal. pitting networks against each other scoring function great.

this worked ttt. not sure how work checkers. chess destroy it. go useless.

the problem neural network learn patters @ fixed location. example jumping opponent in top-left corner totally different situation jumping in bottom left corner. these have learned separately.

perhaps better represent exact state of board in position independent way. require thought. instance might communicate "jump" opportunities exist. move-towards king square opportunity's exist, etc , allow net learn prioritize these.