| Branch | Introduction | Stage | Cost (RTX4090 * Day) | Model size | Notes |
|---|---|---|---|---|---|
| Gom2024 | Gomoku and Renju | Paused | 1500 | b28c512nbt | Sponsored by Zhizi, will continue training months later |
| Hex2024 | Hex game | Finished | 150 | b18c384nbt | See the Release page |
| ConnectSix2024 | Towards solving (proving black win) 19x19 Connect Six | Training | 100 | b18c384nbt | Will finish and release in ~ 5 days |
| Caro2024 | Caro is a modified Gomoku | Finished | 60 | b10c384nbt | https://github.com/hzyhhzy/KataGo/releases/tag/Caro_20240804 |
| LifeGo2024 | Life-and-death Go | Finished | 14 | b18c384nbt b28c512nbt | See the Release page |
| CaptureGo2024 | Capture-Go | Finished | 18 | b18c384nbt b28c512nbt | Will release in 2025.1 |
| ConnectFour2024 | ConnectFour | Finished | 6 | b10c256nbt | See the Release page |
| HexGo2024 | Go on Hex board | Finished | 18 | b18c384nbt b28c512nbt | Will release in ~ 5 days |
| Quoridor2024 | Quoridor | Finished | 7 | b10c128 | Someone has already trained an AI months ago, and found first player win |
| UltimateTictactoe2024 | Ultimate Tic-tac-toe | Finished | 8 | b10c384nbt | See the Release page |
| DawsonChess | 2D Dawson-chess and variants (Reverse Gomoku) | Finished | 1 | b10c384nbt | Chaotic game |
| GoModify1 | A modification of Go | Finished | 6 | b18c384nbt | |
| ProveSevenInARow | See the Readme in the branch | Finished | b10c256nbt | https://mathoverflow.net/questions/471302/can-gomokufive-in-a-row-draw-on-an-infinite-board-what-about-other-m-n-k-game | |
| HexTemplate | Some special openings for Hex | Finished | b20c256 | https://mathoverflow.net/questions/470376/connection-properties-of-a-single-stone-on-an-infinite-hex-board | |
| DotsAndBoxes | - | Finished | b20c256 | Not released yet |
"Black-white No-score 2024" For games which black and white plays alternately
For board games with multi-stage moves or irregular move order For example: Chess(regard one move as two stages: choose and place), Amazons(3 stages), Ataxx(sometimes 1 stage, some times 2 stages), ConnectSix, ...
Latest KataGo engine for Gomoku and Renju. Also called "KataGomo" and "KataGomoku".
Recently a b28c512nbt net is training under the sponsorship of Zhizi(智子)(www.zhizigo.cn www.zhizigomoku.cn), which provides online paid KataGo services.
The training started in 2024.5 and will finish in 2024.11. ~30*RTX4090 for 10 hours every day. The strongest net is only available on Zhizi.
An experiment of combining NNUE and NN. (not very success)
Gomoku-NNUE: A fast and small network with special structure for gomoku.
Gomoku-NNUE is firstly created by me (https://github.com/hzyhhzy/gomoku_nnue/tree/multiRules) and improved by dblue and used in Rapfi(https://github.com/dhbloo/rapfi)
Input the result of a small MCTS (~300 nodes) of NNUE to the big katago net.
Result: The katago net become "as strong as NNUE+MCTS" in a very short time, however further progress is very slow.
If use transfer-training from a trained katago net, it become slightly stronger than the origin net with same nodes, but not worth its large CPU cost.
A branch to test MCTS of NNUE. Regarding the result of a small MCTS (~300 nodes) of NNUE as a "meta-node", then apply MCTS to meta-nodes. Katago NN is not used here. Only to use katago's MCTS
https://en.wikipedia.org/wiki/Hex_(board_game)
Latest KataHex code
Hex2024 with move limit settings, to estimate the length of the optimal game of Hex board on N*N board
Result: f(N) ~ 0.6 * N^1.9
https://mathoverflow.net/questions/302821/length-of-optimal-play-in-hex-as-a-function-of-size/477037#477037
Latest release of Katago for Hex game (KataHex) https://github.com/hzyhhzy/KataGo/releases/tag/Hex_20240812
A special engine for "edge template". Special network is needed.
This branch is to estimate the min width of 8~11th row "edge template" of Hex
Result: f(n)>=9*n-46 for n>=8, and probably f(n)=9*n-46 for n<=11
https://www.hexwiki.net/index.php/Open_problems_about_edge_templates
A special engine for single-stone escape problems (HexTemplate branch is more specific for "edge template"). Special network is needed.
Some analysis and results: https://mathoverflow.net/questions/470376/connection-properties-of-a-single-stone-on-an-infinite-hex-board
Including Caro, ConnectSix, and maybe some math problems
KataGo for Connect Six The 2022 ver has shown that black has great advantage on 21x21 or larger boards, but 19x19 was unknown. Recently @eeoo has found that black has advantage on 19x19, so I started a new run in 2024.8. I decide to strictly prove that black wins in Connect Six on 19x19 boards, collaborating with dblue using his ConnectSix Rapfi.
Caro is a modified Gomoku rule, the only difference is that fives with two terminals blocked are not a win (OXXXXXO is not a win)
Latest release: https://github.com/hzyhhzy/KataGo/releases/tag/Caro_20240804
We all know black wins in 5-in-a-row(Gomoku) on an infinite board. And more than 5 in a row is very difficult which means easy to draw.
Although it's very easy even for a beginner to draw in 6-in-a-row game, 7-in-a-row has not been proven to be a draw in infinite boards and remains to be an open problem for decades, showing a big gap between "theory" and "practice".
https://arxiv.org/abs/2107.05363 This article shows a method to reduce the infinite board to 4*n boards. If we can prove that there exist such one n, that white can draw on a 4*n board (with some kinds of rules), then "7-in-a-row is a draw" is proved. In this article the authors tried n<=14 but failed.
This branch is to solve the game mentioned in this article on 4*n boards with large n.
However katago shows that the min n is around 30, which means it is almost impossible to strictly solve the 4*n game. So this method is almost dead.
More details was written here: https://mathoverflow.net/questions/471302/can-gomokufive-in-a-row-draw-on-an-infinite-board-what-about-other-m-n-k-game
It's a game that black and white alternately play edges. It is hard to represent edges in Katago, so I uses a (2H-1)*(2W-1) board to represent the Dots-and-boxes board Result: This is a very chaotic game, no strategies, just brute-force search. (probably because it is almost an "impartical game"). Katago can learn almost nothing in the first 1/3 game and there is only some strategies near the end of the game
2023.1.8 Engine based on Katago 1.12 start Modifying
| Tasks | Branch | Stage | Notes |
|---|---|---|---|
| 1 Compile and run training | Kata2023 | Finished | - |
| 2 Chinese Rule Only(Remove JP rules) | - | Skipped | |
| 3 CaptureGo(Remove Go rules) | - | Skipped | |
| 4 Black-White board games(Remove Capture) | BW2023/BWnoscore2023 | Finished(outdated) | Using Gomoku as a representative, very easy to be modified to other games. "BWnoscore2023" branch also removed komi and score |
| 4.0 Remove score and komi | BWnoscore2023 | Finished(outdated) | |
| 4.1 Gomoku | Gom2023 | Todo | New Katagomo engine |
| 5 Chess-like games(Multi-stage moves) | Movestone2023 | Finished(outdated) | Use Breakthrough as a representative, very easy to be modified to other games |
| 5.0 Only X-axis symmetry. And flip Y-axis when playing white. (For chess-like games) | MovestoneXsym2023 | Finished | subtreeBias and patternBonus are also removed |
2022.1 Engine based on Katago TensorRT+GraphSearch
Updated on 2022.4.9
Gomoku/Renju: "GomDevVCN" branch
https://github.com/hzyhhzy/KataGo/tree/GomDevVCN
TODO:
先做改动小的，再做改动大的
完全没改的分支：Go2022
| 棋种 | branch | 状态 | 备注 |
|---|---|---|---|
| 1.各种变种围棋（保留中国规则，删掉日韩规则） | CNrule2022 | - | 基础分支 |
| 1.1 加权点目 | weightGo | finished | |
| 1.2 某个子不能死 | aborted | 不做了，因为需要大量死活题作为开局 | |
| 1.3.1 Hex 棋盘的围棋 | HexGo | aborted | 懒得做了 |
| 1.3.2 Hex 棋盘的吃子棋 | HexCaptureGo | finished | |
| 1.4 落子没气先提自己 | firstCaptureMe | finished | |
| 1.5 一子千金 | yiziqianjin | finished | |
| 2.需要提子但不需要点目的 | CaptureGo2022 | - | 基础分支 |
| 2.1 吃子棋 | CaptureGo2022 | finished | |
| 2.2 反吃子棋（不围棋） | CaptureGo2022 | finished | |
| 2.3 活一块就算赢(死活对局) | lifego2 | finished | |
| 2.4 黑棋活一块就赢，但是不能被吃子 | aliveWin | finished | |
| 2.5 吃子 n 子棋 | aborted | 以前做过所以不做了 | |
| 2.6 白能吃黑，黑吃不掉白，黑棋最多多少手不被吃 | EscapeGo | finished | |
| 2.7 谁先没地方下谁输 | yiziqianjin | finished | 视为特殊的一子千金 |
| 3.不需要提子，黑白交替落子的棋 | Gom2022base | - | 基础分支，不跑训练 |
| 3.0 不打算做的 | |||
| 3.0.2 重力四子棋 | aborted | 以前做过，训练了很长时间 | |
| 3.1 五子棋系列(各种规则) | - | ||
| 3.1.0 主线(无禁,有禁,无禁六不胜) | GomDevVCN | training | 100b256f，以后会跑分布式 |
| 3.1.2 禁点五子棋 | GomBanLoc | finished | |
| 3.2 连五的个数 | fivecount | finished | |
| 3.3 反 n 子棋 | AntiGomoku | finished | |
| 3.4 Hex | Hex2022 | finished | |
| 3.4.1 反 Hex | Hex2022 | finished | |
| 3.5 等差数列 6 子棋 | EquDifGomoku | finished | |
| 3.6 Angels and Devils game | AngelProblem | finished | |
| 3.7 每一步必须在上一步的附近某些位置，满足某些条件获胜 | - | ||
| 3.7.1 一种特殊四子棋 | con4type1 | finished | |
| 3.8 Reversi(奥赛罗，翻转棋)，反 Reversi | Reversi2022 | finished | 为了在botzone上打榜，还是做了 |
| 4.一次走两步的棋，或者挪子的棋 | Amazons | - | 把 Amazons 分支作为基础分支 |
| 4.0 国象，中象 | aborted | 别人做过，且效果不好，所以不做 | |
| 4.1 六子棋(Connect6) | Connect6 | training | |
| 4.2 中国跳棋 | tiaoqi | finished | |
| 4.3 Amazons(亚马逊棋) | Amazons | finished | |
| 4.4 Breakthrough | breakthrough | finished | |
| 4.5 Ataxx(同化棋) | Ataxx | finished | |
| 4.6 各种极简象棋变种 | todo | ||
| 5.单人游戏 | - | ||
| 5.1 消消乐 | xiaoxiaole | todo | |
| 6.完全信息，但是有随机性 | todo |
Updated on 2021.4.5 Now Gomoku/Renju code is uploading in "Gomoku" branch.
----------- 2020.7.26 ----------
Now there are gomoku(freestyle,standard,renju,caro), reversi, connect6, breakthrough, hex, "four in a row", Chinese checkers, and many board games few people play.
Now it's the strongest program on this games as I know: all kinds of gomoku(compared with Embryo) connect6(compared with gzero) hex(compared with gzero and) Chinese checkers(compared with Shangxin Tiaoqi)