Train a neural network for tic tac toe with a genetic algorithm

Question

I was trying to implement a genetic algorithm for the game 'tic tac toe'. How I am doing it at the moment is the following:

1. Initiliaze 50 random networks
2. Let each network play against each network.
3. After that, each network will have played 98 games (against each player on each side)
4. the fitness of each network is calculated this way: fitness = wins + draws -loses

5. Apply the genetic algorithm:

   5.1. Select the 5 best networks by using their fitness value.

   5.2. Copy these 5 networks onto the other 45 networks. (So 10 times the first, 10 times the second...)

   5.3. Change p percent of the weights with a random gaussian multiplied with a factor of s.

   if(p < Random(0,1)){ weight/bias += Random.gaussian() * s; }

My values that I am using right now is:

p = 0.05
s = 0.05

And my network is making decisions the following way:

I am using a minimax algorithm with a depth of 2 (not a lot, I know). The evaluation of the board is being done by the network.

My problem right now is that it does not work as expected. When I play against it afterwards, it usually does stupid moves and let's me win or does not win itself when I let it.

I am very happy if someone could help me with this one.

Greetings, Finn