Self-Play Reinforcement Model for Connect 4

A simple CNN that learns Connect-4 through self play.

The model predicts the probability of winning given the current state it is in based on previously played games. That's it, no tree search, no policies.

All the training is through self-play, no external data or players are used.

The two key elements of making this work are exploration (driven here by adding noise to the predictions) and sampling to eliminate correlation between training samples.

Exploration:

The exploration part is developed by adding noise to the neural net predictions:

preds = preds + np.random.normal(0, std_noise, len(preds))

That's it. This leads the model to not replay the same games over and over while also allowing it to play relatively good moves.

Sampling:

Model Architecture:

A simple 3 layer convolutional neural network does the trick here.

Evaluation:

A negamax player:

• checks for a winning move and takes it
• blocks the winning move of an opponent
• avoids moves that allow the opponent to win in the next round

So basically the only way to win against a negamax player is by forcing a win.

Setup:

run python nn_player.py

Play against the model

Gradio Setup

Speeding up the model