A3C Reinforcement Learning for Connect 4

In this repo we provide code to train an AI agent to play Connect 4, and a GUI interface to play against the agent.

Details

This repo implements the Asynchronous Advantage Actor-Critic (A3C) reinforcement learning model to play Connect 4. The board has 7 columns and 6 rows, players take it in turns to drop counters from the top - the winner is the first person to create an unbroken line of four counters (horizontally, vertically, or diagonally).

Useage

To train the agent:

python train_ai_agent.py 8 --steps 5000000

Here 8 is the number of processes, and --steps is an optional parameter to run the training for 5,000,000 training steps.

To use the GUI, one must populate input_path_to_pretrained_agent in GUI.py and then run

python GUI.py

Play is performed by clicking on the top box of the column you want to drop your counter in.

Set-up

This code is adapted from the tic-tac-toe code provided in https://github.com/ruehlef/Physics-Reports In order to run it, one must replace site-packages/chainerrl/experiments/train_agent_async.py with the file train_agent_async.py provided in this repo.