KataGo AI for Ultimate Tic-tac-toe

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:
With Tiebreaker: First player 70%, Second player 14%, Draw 16%
Without Tiebreaker: First player 33%, Second player 11%, Draw 56%

KataGo for Life-and-Death Go

Life-and-Death Go is a variant of Go.
Black has 17~20 initial stones. White's winning condition is having one alive part.

This model is based on official KataGo's b28c512nbt using transfer-learning.

In order to achieve a high level of Go strength at a very low cost, the early stages of training were done on a small 9x9 board, followed by 13x13 board, and finally 19x19 board.
Additionally, the official KataGo weights were used for transfer learning on the "life-and-death" game rules, which significantly improved the model's strength.
~ 1 week on 2x RTX4090

KataGo AI for ConnectFour (Four in a row)

https://en.wikipedia.org/wiki/Connect_Four
This version supports all size of boards <= 20x20
This version has much fewer wrong evaluations than the old version(20210511).

The following chart is the results of every board size <= 20x20 calculated by KataGo
Although these results are very sure, the possibility of errors cannot be completely ruled out.

KataHex 20240908: Move number limit mode

The only update of this version is move number limit mode
Some analysis and results of the optimal game length of Hex

This version is slightly weaker (-50~100elo) than 20240813 version.
If you don't need move number limit function and want it to be as strong as possible, you can download 20240813 version

There are two packages.
OpenCL only: Compatible for all GPUs. But slower than TensorRT on Nvidia GPUs
TensorRT+Cuda+OpenCL: For Nvidia GPUs only. Include three types of engines. TensorRT is the fastest but only available for Nvidia GPUs. This package include many DLLs of TensorRT and Cuda, so this package is much bigger.
"raw_data_and_scripts_of_optimal_game_length_analysis.zip" is some scripts to create figures, which is useless for users.

Usage: Press "E" in the LizzieYzy GUI , then enter the command: mm [number]
mm [number] Limit the move number. When reaching the move number without anyone win, the game will be regarded as a draw
mm 50 Limit to 50 moves
mm 0 Disable move number limit
For boards larger than 27x27, this function is very inaccurate.

KataHex 20240812

This is a new version of KataHex (Katago for Hex game)
Using newer model structure (b18c384nbt) and with much more training, the new version becomes much stronger.

Running selfplay on large boards is too slow(~O(n^4) to O(n^7)). So I started training from a smaller board 15x15 (the 2022 version started from 13x13). At the final stage I continue training the model on larger boards, called "Transfer learning".
This version was trained on 15x15 board using 2*RTX4090 for 2 month, ~40M selfplay games.
Then on 19x19 for 15 days, ~2.6M selfplay games.
Finally 27x27 for 3 days, 90k selfplay games.

Old version: https://github.com/hzyhhzy/KataGo/releases/tag/20240406 (Using b40c256 net trained in 2022)
New vs Old, 500 games with balanced one-move openings, 100 visits per move.
13x13 board: +56 elo (290:210)
14x14 board: +273 elo (414:86)
15x15 board: +268 elo (412:88)
19x19 board: +263 elo (410:90)
25x25 board: +466 elo (468:32)
33x33 board: +511 elo (475:25)
Testing on 11x11 board is meaningless because it was almost solved even with any one-move openings.
The largest board which KataHex can play well is around 33x33 ~ 41x41.

There are two packages.
OpenCL only: Compatible for all GPUs. But slower than TensorRT on Nvidia GPUs
TensorRT+Cuda+OpenCL: For Nvidia GPUs only. Include three types of engines. TensorRT is the fastest but only available for Nvidia GPUs. This package include many DLLs of TensorRT and Cuda, so this package is much bigger.

Katagomo-Caro 20240804

Update on 2024.8.4 : Fix a serious bug of SIXWINRULE_NEVER

This is a special version of Katagomo for Caro rule
Trained on 2*RTX4090 for 1 month
Caro is a modified Gomoku rule, the only difference is that fives with two terminals blocked are not a win. For example, OXXXXXO is not a win

Packages for Gomoku and various games

Not new packages. Just uploading some old packets on github for convenience.
If you are in China and downloading on github is very slow. There's a qq group: 1049389629
Sorry there is no English instruction in the packages, only Chinese.
To avoid massive dlls of CUDA, there is only OpenCL version(except Gomoku). If you need TensorRT engine(2x speed on RTX series), please compile yourself

Game list and corresponding branch of each package:
Gomoku(Gomoku,Renju)--Gom2024
Connect6--Connect6
Hex--Hex2022
Ataxx--Ataxx2023
TiaoQi(Chinese checkers)--tiaoqi
MoveStoneGames(Amazons,Ataxx,Breakthrough,TiaoQi)--Amazons,Ataxx,breakthrough,tiaoqi
ConnectFour-connectfour

New packages

New packages with graphic interface.
Strongly suggest using this instead of Gomocup version.
This is 300x~1500x faster on RTX GPUs than gomocup(single thread CPU) version. +1000~1500 elo
Use Lizzie-yzy interface.

Katagomo_Lizzieyzy_Package_20210520.7z is basic package(Opencl Engine)
If you want CUDA engines or CPU-only engines, replace "katago" folder with the one in additional packages.

Connect Four(重力四子棋)

Connect Four Katago with graphic interface.
有图形界面的重力四子棋ai
某大作业的一部分内容

New package with stronger networks

Sorry there is still only Chinese instructions.