AlphaZero for mere mortals: self-play Gomoku on one PC

What it does

This project is a pure self-play implementation of the AlphaZero algorithm targeting Gomoku, also known as Five in a Row. Because the game is far less complex than Go or chess, the training pipeline collapses to something a single PC can handle. You end up with a neural network guided by Monte Carlo Tree Search that learns board strategy from nothing but its own games.

The interesting bit

The cleverness is in the shrink-wrap: the author uses Gomoku’s simplicity as a test tube for the AlphaZero training scheme itself. It is less about building the world’s best Gomoku engine and more about making the algorithm tangible and tweakable on commodity hardware.

Key highlights

• Trains from pure self-play with no human game data or heuristics.
• Supports three deep-learning backends: Theano/Lasagne, PyTorch, and TensorFlow.
• A reasonably good 6×6 model trains in roughly two hours; an 8×8 board needs closer to two days on a single PC.
• Ships with four pre-trained models so you can play against the AI immediately.
• The network logic is isolated in policy_value_net.py, making it straightforward to port to another framework.

Caveats

• The default training path still points to Theano and Lasagne, which now require version-pinning workarounds.
• Switching to PyTorch or TensorFlow currently requires swapping imports inside train.py.
• The bundled pre-trained models were built with Theano/Lasagne, so using them in PyTorch requires extra conversion steps.

Verdict

Grab this if you want a compact, hackable AlphaZero reference you can actually train at home. Skip it if you are looking for a polished, framework-agnostic engine or state-of-the-art Gomoku performance.