This project demonstrates a simple example hot to train neural networks that can play simple games like TicTacToe and Connect 4.
Clone repository and change to the cloned source dir.
git clone https://github.com/macoun/simple-ai-games.git
cd simple-ai-games
Create and source a virtual env. (Example for macos/linux)
python3 -m venv --prompt venv venv
source venv/bin/activate
Install requirements.
pip install -r requirements.txt
The entry point is connect4.py.
python -m connect4
This will start a gui game with a trained AI (neural network) player.
To play the game in your terminal, use the --mode console option.
python -m connect4 --mode console
The entry point is tictactoe.py.
python -m tictactoe
This will start a gui game with a trained AI (neural network) player.
To play the gamein your terminal, use the --mode console option.
python -m connect4 --mode console
The following configurations applies for both, connect4 and tictactore games.
To play against a NN player, you can use the option --player nn.
python -m connect4 --player nn
Since this is the default player, you don't need to specify it explicitly.
The NN player is trained with a default number (10000) of simulated plays.
To change the simulated play count, use the --simulations option.
python -m connect4 --simulations 40000
or similarly for tictactoe
python -m tictactoe --simulations 40000
To change the default number of epochs for training, use the --epochs option.
python -m connect4 --simulations 40000 --epochs 2
The above example can take a couple of minutes to train, but the result is a pretty challanging Connect 4 game.
The minimax player uses the minimax algorithm to find the next best moves by traversing a tree of game states.
To play against a minimax player, use the option --player minimax.
python -m connect4 --player minimax
To specify how deep to traverse the tree before deciding
which move to choose, use the --lookahead option.
python -m connect4 --player minimax --lookahead 5
The minimax players can play almost perfect games if you
set the lookahead high enough.
You can simulate multiple games with NNPlayer vs. MiniMaxPlayer
players to check how well your model is playing against
a challenging opponent.
You will quickly notice that using minimax will take some time, especially with connect4.
We could accelerate the algorithm by applying alpha/beta pruning to some branches in the state tree, but it won't make this kind of games significantly faster if we don't use heuristics. Otherwise, we need to build the entire possible game tree down to the last leaf nodes (game over state) in order to identify the winner.
... or we can use a NN player ;)