I've created a neural network trained by genetic in C++ for the game reversi (also known as othello). The neural network can run... but its intelligence doesn't seem to increase even after tens of thousands of generation.

I believe the problem lies in how I determine the fitness for each genome. What I do is:

pop[0].dFitness = 0;
for (i=1;i<popSize;++i){
bFitness = 0;
int bj = 0;
for (j=0;j<i;++j){ // Find the current best genome
if (pop[j].dFitness > bFitness){
bFitness = pop[j].dFitness;
bj = j;
}
}
NN2.PutWeights(pop[bj].vecWeights);
NN.PutWeights(pop[i].vecWeights);
Play(p1,p2); // match current genome against current best
pop[i].dFitness = p2 - p1 + 64;
}

So, the fitness score for each generation differs because the fitness score is relative to the best in the current population. This means that when a population evolve, genomes may not evolve into something smarter; instead, they may evolve into a network which possibly can win the previous best, but may actually be a weak network. Basically it goes kind of a loop, like rock scissors paper.

If you have any idea how i can solve this problem, please do share! Thanks :)

I believe the problem is with your last equation, and I'll give the reasons why I think it is. Not completely familiar with the problem and all that it entails, so I hope this doesn't seem off in left field.

Given 64 pieces on the board, I can understand why initially that the equation would be p2-p1 + 64... this would give you the positive value difference between the two, but this is assuming that 1 player completely dominating the other is the most positive common outcome.... Your case assumes that 64-0 + 64 = 128 is the objective, but it's the most difficult to achieve.

Take this equation (p2-p1 + 64)/(p2-p1)
for (64,0) your outcome = 2
for (48,18) your outcome = 3.133
for (40,24) your outcome = 5
for (33,31) your outcome = 33
for (31,33) your outcome = -31
for (18,48) your outcome = -1
for (1,63) your outcome = -.032
(note: using this equation you would need check for (32,32) and replace the equation with the value of a constant, which value I'm not sure of)

The closer you get to evenly matched, the higher the score... because of the higher probability of that outcome. The problem with this given equation is that (31,33) results in such a large negative value, and the worse you do, you get better. I would guess your looking for a 3rd (possibly 5th degree), two variable, approximation equation.

Graphically (values given are p2,p1... not points on the graph), something that starts mid-range below the origin at (0,64), reaches a min at (18,48), reaches a horizontal (not vertical) inflection point at (32,32) (and crosses the horizontal axis), reaches a max at (48,18), and finishes mid-range high at (64,0). An S fallen forward.

For an exact equation... Some how, at least 2 more values should enter the equation... 60! (60* 59* 58...) (on the top) and 64! (on the bottom)??? 60 because you start wiht 4 on the board, so you only have control of 60 pieces. Could be only 30!, noting the number of decisions 1 person has to make. May need to also factor in the limited number of positions that you can actually place a piece each turn, which starts with 12, then (12 - 1 + 3)... but this would only last until you reached the edge of the board... I'd guess the average is 30^12, noting that as the end of the game is reached you have less and less available choices (not necessarily positive choices), so it should even out. How it all fits together, I don't really know off hand (I should have paid better attention in college). If you figure it out, it would be interesting to know.

Tip: find a college statistical analysis book (preferably engineering)... the equations are in there some where.

Good luck,
Gerry