viliamvadocz / tak

An implementation of AlphaZero for the board game Tak

AlphaTak

About

An implementation of AlphaZero for the abstract strategy board game Tak.

The goal of this project is to advance the state of the art in Tak bots, increase the skill level of the game, and to experiment with the Rust Machine Learning ecosystem.

I used tch-rs crate (wrapper for libtorch) as my machine learning library. I experimented with tensorflow as well, but I found the documentation lacking in many aspects. Overall, the quality of resources for machine learning in Rust is very poor.

Building

It is strongly recommended that you manually install the C++ library PyTorch (libtorch) v1.11.0. The instructions for installing can be found here.

Then you can just run cargo build --release in the tak directory. This will build all the packages in the workspace. If you want to build a specific one instead, you can specify like this: cargo build --release -p <package-name>.

Brief summary of what each package does:

• tak library: implementation of Tak including move generation and parsing of PTN and TPS
• alpha-tak library: implementation of AlphaZero, MCTS, the network
• train binary: training the network with self-play
• analysis binary: interactive local analysis
• playtak binary: for running the bot on playtak

tak's Issues

Racy file write

tak/playtak/src/main.rs

Lines 90 to 93 in c3f0fe9

	// create analysis file
	if let Ok(mut file) = File::create(format!("analysis_{}.ptn", sys_time())) {
	file.write_all(player.get_analysis().to_ptn().as_bytes()).unwrap();
	}

These lines are not guaranteed to run -- once the main thread ends, all background threads are eligible to be killed immediately.