Reinforcement learning with Pytorch

تفاصيل العمل

What you are looking at is my most recent project

in which I trained an AI agent to play this simple board game.

what makes this project very special is that the AI is seeing the game just the same way you are looking at "a picture".

so the AI model has to train its vision as well as training its strategy to play the game, to seek the black cherry and avoid all poisions

due to this, it took me many hours of training and intense processing load and that's why I used my graphics card to train it.

although it's not perfect yet and it loses sometimes, it's just a matter of time to master the game and win e each round.

I used Pytorch

and the following is the structure of the value neural network:

Conv2d(in_channels=1, out_channels=16, kernel_size=6, stride=1, padding=2)

Conv2d(in_channels=16, out_channels=4, kernel_size=4, stride=3, padding=2)

the decision is added to the output of the previous layer and then transfered to the following layers

LeakyRelu(150 units)

LeakyRelu(45 units)

Sigmoid(2 units)