This repo holds code for a multi-agent RL model for playing connect4 that is trained with curriculum learning
I trained this model on the Connect4 PettingZoo out of the box environment (found here)
I trained it using a DQN developed in AgileRL (found here)
The first curriculum I trained the model on (can be found as lesson1.yml). Here I train the model against a randomly-picking opponent (10,000 games)
The model's evaluation significantly increased over training, but did not converge
In the second curriculum (can be found as lesson2.yml). I trained the model to play against itself (50,000 games)
The model performed very well in evaluation (playing against randomly playing opponent)
Here's an example of the current iteration of the agent (red) beating a randomly picking agent (black)
And an example of the agent playing against itself
