katagomo - KataPente (KataGo for Pente) 20250726

Pente rule

https://en.wikipedia.org/wiki/Pente
This version include Pente and Pente-Keryo

Training information

About 8 RTX4090-days, 270M selfplay games. The mainline training was using SGD optimizer. After mainline finished a b10c256nbt was retrained using Muon optimizer and 50 elo stronger than the previous model.

Scripts used in this run: https://github.com/hzyhhzy/KataGomo/tree/Pente2025/scripts

Results

More detailed results are in KataGomo-Pente.Introduction.pdf Almost every opening of Pente is easy enough to be solved within <10M search nodes (a few minutes with a good GPU). So even swap rules (pie rules) cannot make the game fair.

First player wins in 35 moves

Results of 3-moves openings

Published by hzyhhzy 7 months ago

katagomo - KataGo Ultimate-Tic-tac-toe 20250616

Update

This version fix a serious bug (the nn can't see the previous move) (Thanks for @Daporan for finding it) and retrained.
Strength +100 elo

Rule

https://en.wikipedia.org/wiki/Ultimate_tic-tac-toe

This AI includes a widely-accepted additional rule: If one should play in a finished (win, lose, or filled) sub-board, then he/she can play everywhere. It has been proved as first player win without this rule.

This AI supports both with and without “Tiebreaker” rule. Tiebreaker: If no one connects 3 subboards, then who own more subboards wins.

Results:

The following images shows the winrates(up) and drawrates(down) of the first move With Tiebreaker: First player 78%, Second player 4%, Draw 18%

Without Tiebreaker: First player 21%, Second player 2%, Draw 77%

Published by hzyhhzy 8 months ago

katagomo - Kata-Breakthrough 20250616

Rule of breakthrough

https://en.wikipedia.org/wiki/Breakthrough(boardgame)

Usage

Download Kata-Breakthrough_20250616.7z and read ReadMe.pdf
Notice the interface cannot show the game board correctly.

Training

Training scripts used in this run:
https://github.com/hzyhhzy/KataGomo/tree/Breakthrough2025/scripts

b18c384nbt model. 22M parameters
~8M games on <= 8x8 board
~2M games on <= 10x10 board

Winrates of Breakthrough of <=10 boards

Published by hzyhhzy 8 months ago

katagomo - KataGomo Gomocup version

Tiny b10c128nbt (1/50 size of full strength model b28c512nbt) models for Gomoku/Renju/Caro.
Maybe 400 ~ 600 ELO weaker (not tested) than full strength version with the same number of search nodes.

The engines attached here are single-thread cpu-only version ( ~40 v/s) which is much slower than GPU versions (5000 v/s for a 50x size net b28c512nbt on an RTX4090).
Full strength GPU version (1000 ~ 1500 ELO stronger) is here https://github.com/hzyhhzy/KataGomo/releases/tag/Gomoku_20250206.

However Gomocup does not allow submitting different engines for different rules. So finally I decide to quit it this year.

Published by hzyhhzy 9 months ago

katagomo - KataGo AI for Ultimate Tic-tac-toe 20241101

Update

The only update is stronger net +50 elo
The old version(20241019) has a serious bug of training openings. New version fixed it

Rule

https://en.wikipedia.org/wiki/Ultimate_tic-tac-toe

This AI includes a widely-accepted additional rule: If one should play in a finished (win, lose, or filled) sub-board, then he/she can play everywhere. It has been proved as first player win without this rule.

This AI supports both with and without “Tiebreaker” rule. Tiebreaker: If no one connects 3 subboards, then who own more subboards wins.

Results:

The following images shows the winrates(up) and drawrates(down) of the first move With Tiebreaker: First player 70%, Second player 14%, Draw 16%

Without Tiebreaker: First player 33%, Second player 11%, Draw 56%

Published by hzyhhzy 10 months ago

katagomo - KataGo for Score Four (and five, six, eight) 20250510

Introduction

Rule of Score Four : Connect Four with gravity on 4x4x4 3D board.
https://en.wikipedia.org/wiki/Score_four

“Score N”: Connect N with gravity on NxNxN 3D board.
This AI includes N=4,5,6,8

Source: https://github.com/hzyhhzy/KataGomo/tree/ConnectFour3d

Results

Score Four : Probably the first player wins Conjecture of Score N with N>=5: The first player wins if N is odd; The second player wins if N is even;

There is no technical difficulty in solving Score Four strictly (by combining KataGo with a solver like https://tum-csc.github.io/c43d/) , but it is quite cumbersome. So I'm not planning to do it now. Hope someone will do it.
More detailed results can be found in ReadMe.pdf

Training

b10c256nbt network structure
Scripts: https://github.com/hzyhhzy/KataGomo/tree/ConnectFour3d/scripts

Published by hzyhhzy 10 months ago

katagomo - KataGo Capture-Go 20250509

Capture Go: Who first capture N stones wins. Pass is not allowed so the game will finish when one player is forced to suicide.

This released package support board size ≤19, capture target N≤5
This released package has only OpenCL engine. If you need TensorRT engine, compile by yourself.
The network is b18c384nbt using "transfer learning" from official KataGo

Press "E" in the interface and enter cap N to set capture target .

Example game opening of cap 1 on 19x19 board

Published by hzyhhzy 10 months ago

katagomo - KataGo Connect6 (Connect Six) 20250505

File list

KatagomoConnect6_20250505.7z Full package. Include TensorRT engines which is faster than OpenCL for Nvidia GPUs.
KatagomoConnect6_20250505_OpenCL-only.7z OpenCL-Only package. Include all features and much smaller, just slower.
Connect6_19x19.MoveLimit113.Openings.7z An opening lib (HTML format) with move limit 113 (if no one wins before 113 stones, then draw). And regard draws as black loses.
connectsix19x_b18trans.bin.gz Model for boardsize <= 19x19 connectsix25x_b18trans.bin.gz Model for boardsize <= 25x25

Results

19x19

Black has a significant advantage. Opening library: Connect6_19x19.MoveLimit113.Openings.7z Estimated minimum stones needed for a forced Black win: 109 (109 stones = 55 moves = 28 rounds).

18x18 or smaller

Highly likely to end in a draw

20x20

Insufficient computational analysis completed. No significant advantage for Black currently observed.

21x21

Black has a stronger advantage. Estimated stones needed for a forced Black win: 93 (93 stones = 47 moves = 24 rounds).

25x25

Black's advantage increases further. Estimated stones needed for a forced Black win: 89 (89 stones = 45 moves = 23 rounds).

Published by hzyhhzy 10 months ago

katagomo - KataGo Animal chess (Dandelion) v2.4

https://github.com/lxsgx23/Dandelion-Chess/releases/tag/v2.4 Stronger nets, better interface

Published by hzyhhzy 10 months ago

katagomo - KataGo Zhenqi 20250218

“震棋”是“逆界黑白”群发明的一种新棋

震棋规则：双方轮流落子，落子后会击退周围一圈所有棋子1格，如果击退到的格子已经有棋子了则不能击退，可将棋子击退出棋盘外移除掉。落子后或者击退棋子后棋子形成四连立刻获胜。 规则补充：落子后先判定是否连四，然后再击退，然后再判定击退后的局面是否连四。若击退后双方同时连四，则走棋方胜。连五及以上也算连四

结论：6x6以及较小的棋盘很可能是和棋 8x8棋盘黑棋疑似必胜，且双方都走最优解时黑棋取胜步数为41 10x10棋盘黑棋疑似必胜，且双方都走最优解时黑棋取胜步数为35

Published by hzyhhzy about 1 year ago

katagomo - KataGo-Nogo 20250219

Nogo rule: The capture rule is the same as Go. Who firstly capture one stone loses.

b28c512nbt model ~9M games on 9x9 board ~0.5M games on 13x13 board

Result

Published by hzyhhzy about 1 year ago

katagomo - KataGo 斗兽棋（Animal Chess, DouShouQi）+ Dandelion 界面

Interface: Dandelion v2.2 Source code: https://github.com/lxsgx23/Dandelion-Chess Authors: Laoxu(Kouza) QQ:3527355284 github.com/lxsgx23 Sigmoid(hzyhhzy) QQ:2658628026 github.com/hzyhhzy

Engine: KataGo-AnimalChess Source code: https://github.com/hzyhhzy/KataGomo/tree/AnimalChess2025 Authors: Sigmoid(hzyhhzy)

Conclusion: Almost draw (draw rate >99%) Even with leopard+wolf+dog odds, draw rate is still >60%

Published by hzyhhzy about 1 year ago

katagomo - KataGomo 20250206: Stronger Renju/Freestyle/Standard/Caro model and Zhizi cloud

KataGomo(KataGo-Gomoku, 五子棋KataGo) 20250206

Stronger b28c512nbt model sponsered by Zhizi 更强的b28c512nbt模型，“智子围棋”赞助训练。

Include Renju/Freestyle/Standard/Caro four rules and Zhizi cloud-computing 内置Renju/Freestyle/Standard/Caro四种规则，以及智子云租机

For more details, see the introduction file attached. 更多细节见说明文档

2025-02-05 21:12 UTC : Fix error: add msvcp dlls 修复：添加msvcp的dll

2025-02-08 19:27 UTC : Fix error: Zhizi engine sometimes does not work 修复：智子引擎有时候无法正常使用

QQ: 2658628026 QQ Group: 1049389629 Discord: hzy_sigmoid

Published by hzyhhzy about 1 year ago

katagomo - KataAtaxx 20250131: Stronger model, TensorRT engine

Stronger b18 model: +100 elo TensorRT engine: NVIDIA gpus only, ~2x speed of OpenCL engine

'setfen' and 'getfen' GTP command to handle FEN string able to set blocked locations

Published by hzyhhzy about 1 year ago

katagomo - KataHex 20250131: Stronger b28c512nbt model

Retrained b28 model using the data of 2024 version. +100 elo on 15x15 +200 elo on 19x19

hex3_27x_b28.bin.gz Strongest model for boardsize <= 27x27 , does not support move-limit mode hex3_mm19x_b28.bin.gz Support move-limit mode but slightly weaker.

You can download the new package KataHex_20250131.7z. If you have downloaded the 20240908 version, you can also replace the model of the old package in ./weights. The engine has no update Last release: https://github.com/hzyhhzy/KataGomofork/releases/tag/Hex20240908

Published by hzyhhzy about 1 year ago

katagomo - See the old repo

https://github.com/hzyhhzy/KataGomo_fork/releases

Published by hzyhhzy about 1 year ago