Companion Baduk

Hello everyone!

I wanted to share a project I’ve been working on for the past few years called Companion Baduk, which I recently released on Steam!

I originally started developing the application to help beginners learn how to play Go. My goal was to make learning the game a little more fun, accessible, and engaging through gamification, while still providing useful tools for players who want to improve.

Over time, the project has grown quite a bit, with different AI opponents, game modes, puzzles, lessons, and analysis tools.

One of the features I’ve spent the most time developing is the game analysis and review system. It uses an AI system I’ve been working on that combines existing Go engines with a dictionary of recognizable shapes, patterns, and proverbs. Rather than simply recommending the strongest move according to an engine, it focuses on finding moves that can be explained in a way that’s useful for learning.

The idea is to help players understand why a move might be good, with explanations based on things like shape, liberties, connections, captures, and life-and-death patterns. You can explore different positions during a game review and get live recommendations and explanations as you go.

The AI can also be played against directly, with different difficulty levels based on how it selects moves from its dictionary of Go shapes, patterns, and proverbs:

  • Around 30 kyu: Randomly selects moves from recognized shapes and patterns.

  • Around 20 kyu: Prioritizes the most commonly occurring shapes and patterns.

  • Around 1 dan: Uses KataGo to evaluate the available shape and proverb-based moves, selecting the strongest option.

The game analysis system always uses the strongest version, which plays at approximately 1 dan strength. This allows it to provide reasonably strong recommendations while keeping its suggested moves tied to recognizable Go concepts that can be explained to the player.

I’ve found this to be an interesting balance between playing strength and teaching. Even at its highest difficulty, the AI is restricted to moves it can explain through recognizable Go concepts. I’m pretty happy with how well it performs, especially since the goal isn’t necessarily to find the absolute best move, but to recommend good moves and help players understand why they work, rather than just telling them where to play.

The application itself has an anime-inspired art style, with different AI companions who have their own personalities and playing strengths. There’s also a progression system, interactive lessons, puzzles, and various ways to practice.


I’ve also included quite a few different game modes to keep things interesting, including:

  • Normal Go, with different board sizes and settings.

  • Handicap Go, for playing against stronger opponents.

  • Chinese and Sunjang Openings, featuring traditional starting arrangements.

  • Tiny Board Gauntlet, where you face a much stronger AI on a 5×5 or 7×7 board.

  • High Komi Comeback, with unusually high komi to shake up the game.

  • Same-Color Showdown, where both players’ stones appear the same colour.

  • Move Timer Blitz, with just 20 seconds to make each move.

  • Chaos Board Challenge, where the board periodically rotates or mirrors itself.

  • Atari Start Skirmish, starting from a small board with stones already fighting.

  • 17×17 Scramble, for something a little different from standard board sizes.

There’s also a Mystery Challenge mode that randomly selects one of these special challenges, so you never quite know what kind of game you’re getting into!

There are quite a few other features as well, including ranked puzzles, joseki practice, and a collection of Go proverbs.

I’ve also spent a lot of time working on the proverb database and integrating it into gameplay. You can freely explore the collection, but what’s more interesting is that your AI companion can recommend relevant proverbs based on the current board position, both while playing and during game reviews. The idea is to help players connect traditional Go wisdom with actual situations they encounter in their games.



It’s been a pretty big personal project, combining my interest in Go with what I’ve learned studying computer science. I’ve particularly enjoyed developing the AI and experimenting with ways to make Go concepts easier to understand.
I’m still actively working on the game and adding new features, particularly improving the AI explanations and recommendations to make them more useful for beginners.

AI Engines and Game Server

Another big part of the project has been experimenting with different Go AI engines. I’ve tried to find and integrate as many freely available Go engines as possible, each with their own playing styles and strengths.

To help test and compare them, I developed a Go server that automatically runs games between different AI engines. The server adjusts the games using 2–5 handicap stones to try to maintain roughly a 50/50 win rate between opponents, allowing engines of different strengths to compete against each other.

It’s been really interesting watching these games and using different tools to track how the engines perform over time. I’ve also been developing my own AI and having it compete against the other engines on the server, which has been a fun way to test improvements and see how well different approaches work in practice.

The server eventually became one of the main backend systems for Companion Baduk, handling communication between engines, running matches, and managing games.

It’s probably been one of my favourite parts of the project, especially getting to experiment with so many different AI systems and see how they play against each other.

If anyone is interested, you can check it out here:

Steam: Save 15% on Companion Baduk on Steam

Website: Companion Baduk — Your next Go adventure.

Feedback Form: Beginner GO Game feedback

I’d really appreciate any feedback! I’m particularly interested in hearing what people think about the explainable AI approach to game reviews, whether the explanations are actually helpful, and what could be improved.

Some additional screenshots and technical details

AI proverb and shape weights

These graphs show the weights assigned to different Go proverbs and patterns under various playing conditions. I’ve been experimenting with how the importance of different concepts changes depending on the board position and whether the AI is playing an even or handicap game.

Even games on 19×19:

Playing White in 19×19 handicap games:

Playing Black in 19×19 handicap games:

These are used to help curate the review section and proverbs that get recommended to you during the game.

Steam Deck and smaller screens

I’ve also been working on making the interface more comfortable to use on smaller displays, including the Steam Deck.

Distribution of the different AI on the application

Assorted stats from the application

Puzzles generated from player mistakes

Another feature I’ve been experimenting with is automatically generating puzzles from mistakes made during games. The idea is that when a player makes a significant mistake, the position can be saved and turned into a puzzle for future practice.

This way, players can gradually build their understanding by practicing positions based on mistakes they’ve actually made.

Thanks for checking it out! I’m always happy to hear suggestions or discuss any of the technical details with anyone who’s interested.

Other Go Projects

Along the way, I’ve also worked on a few other Go-related projects that were originally developed for Companion Baduk but eventually got set aside or turned into their own applications.

GoGuesser: https://goguesser.onrender.com/

A multiplayer Go game where players work together to analyze board positions, discuss possible moves, and vote on what they think the best next move is. It’s open source and was an experiment in making Go analysis a more collaborative experience.

GoLibrary: GitHub - axyl-casc/GoLibrary · GitHub

A Go library application I developed to make it easier to organize and browse Go resources in one place, including SGF game records, PDFs, and HTML files.

Both started as ideas for Companion Baduk, but I ended up developing them separately as I experimented with different ways to make learning and studying Go more accessible.

6 Likes

I love that idea.

2 Likes

It’s a very impressive amount of work.

I’m interested in your bot design, if you don’t mind me asking. So you use a for example 1d Katago bot and limit it’s move selection to moves that fit the (which are in the legend of the graphs).

I assume you do some benchmarking to develop the weights. Have you benchmarked this 1d-limited-by-proverbs bot vs. 1d-normal bot? I assume the proverbbot is a bit weaker, but I wonder how much stones the difference is.

Also, do you automate the weight training or do you do it by hand?

1 Like

Hi, thank you!

I use a smaller KataGo network (kata1-b6c96-s175395328-d26788732.txt.gz) to evaluate moves generated from a database of Go shapes, patterns, and proverbs. Each candidate move is evaluated for its win rate, and the highest-rated move is selected by the teaching AI.

This system plays at roughly a 1d level, slightly stronger than the human-like 1d KataGo bot, but weaker than the unrestricted KataGo network. For comparison, the unrestricted network plays at around 3d, so limiting it to proverb-based moves makes it approximately 2 stones weaker. I felt this was more than strong enough for teaching newer Go players and letting them explore different moves and understand why they might be good.

The weight training is all automated within Companion Baduk itself!

I also have a few beginner-level AIs that use the collected training data to build models of commonly played shapes and patterns. These bots then semi-randomly select moves based on those models, giving me a range of beginner opponents from around 40k to 20k.

This works out nicely because the weakest KataGo human-like AI is around 20k, so I can offer a fairly consistent progression of AI opponents throughout the program, from complete beginners to stronger players.

3 Likes

This is a really impressive project! I especially like the idea of using AI to explain why a particular move is good instead of simply suggesting the strongest move. That kind of feedback can make learning Go much easier for beginners. Turning mistakes from actual games into practice puzzles is another great feature. It’s clear you’ve put a lot of thought and effort into making the learning experience more engaging. Great work!

Thank you for the encouraging words! It’s been a lot of trial and error getting everything to work together. The lessons and puzzles are still a work in progress, but I’m pretty happy with how it feels to play against the AI now.

I’ve just released another update with improvements to the AI, puzzles, and game reviews, along with Steam achievements! I’ve also been working on expanding the companions’ backstories and the world they inhabit, trying to make the experience a little more immersive.

There’s still plenty to improve, but I’m happy with how far it’s come!

For anyone interested, you can read about the latest update here: