Why does a neural network stop learning?

Question

Python

I am writing a neural network that would learn to play the game "4 in a row". Code
Used convolutional neural network and q-learning. Another difficulty is that you cannot change the Q function after each move. You need to remember the move, and only after the game has been played, change the Q-function.
But oddities arise, because after 1-3 epochs, all moves become the same.
This is proven by the evaluate() function
Logs:

Epoch 0/10000
Right ANS=410, Real=[8, 8, 8]
Epoch 0/10000
Right ANS=410, Real=[8, 8, 8]
Epoch 100/10000
Right ANS=410, Real=[8, 8, 8]
...
Epoch 9900/10000
Right ANS=410, Real=[8, 8, 8]

The neural network itself:
InputConv2d (1 filter)
Conv2d(32 filters, kernel size = 4x4, because the game is called 4 in a row)
perceptron 60
perceptron 30
perceptron 10 - output
What is the problem? (I feel like no one cares. Or go figure it out yourself, why should I waste time on you)