KataHex - Revision history

Hexanna: mention the 20250131 net; fix links

2026-04-18T12:47:38Z

mention the 20250131 net; fix links

Selinger: /* Links */ Added newest pre-trained networks

2025-01-31T17:41:28Z

‎Links: Added newest pre-trained networks

Selinger: /* Links */ Updated github links to Hex2024 branch.

2024-11-01T00:38:16Z

‎Links: Updated github links to Hex2024 branch.

Selinger: /* Links */ Removed link to Chinese website that has disappeared

2024-10-31T00:17:39Z

‎Links: Removed link to Chinese website that has disappeared

Selinger: /* Links */ Provided a direct download link for the 20240812 model

2024-10-31T00:15:55Z

‎Links: Provided a direct download link for the 20240812 model

Hexanna: the new net does better in arbitrary positions

2024-08-17T21:34:00Z

the new net does better in arbitrary positions

Hexanna: new katahex release

2024-08-13T03:12:17Z

new katahex release

Selinger: /* Pre-trained networks */ Clarification

2024-01-16T23:49:10Z

‎Pre-trained networks: Clarification

Selinger: /* Pre-trained networks */ Fixed link syntax.

2024-01-11T00:37:10Z

‎Pre-trained networks: Fixed link syntax.

Selinger: New page

2023-11-18T14:38:06Z

New page

New page

'''KataHex''' is a free and open-source computer Hex program, capable of defeating top-level human players. It implements Monte Carlo tree search with a convolutional neural network providing position evaluation and policy guidance.

== History and versions ==

KataHex is based on [https://en.wikipedia.org/wiki/KataGo KataGo], a computer Go program developed by David Wu that was first released on 27 February 2019. It was adapted for Hex by "HZY" between February 2020 and May 2022. While initially unnamed, the Hex-adaptation of KataGo quickly became known as "KataHex" among Hex players.

The HZY implementation of KataHex speaks a non-standard dialect of [[GTP]] and can only interact with a specially modified version of the Go GUI known as LizzieYzy. A further adaptation of KataHex that is capable of interfacing with [[Hexgui]] was made by [[User:Selinger|Selinger]].

== Pre-trained networks ==

HZY initially trained the neural network model on two NVIDIA GeForce RTX 2080 Ti GPUs for about 20 days. They then up-trained the network on 19x19 for one day, and on 27x27 for an additional day. [[https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en|According to HZY]], the up-trained 19x19 network is relatively reliable, but the 27x27-network is not.

Each pre-trained neural network is able to play Hex at the size it was trained on, or any smaller size. Thus, the 19x19 network model can also be used to play 11x11 or 13x13.

"KataHex" refers generally to the KataHex program, but often more specifically to the pre-trained 19x19 neural network made by HZY.

== Limitations ==

Since KataHex was only trained on self-play, it does not always do well when asked to play from an arbitrary starting position. It seems to be particularly thrown off if the starting position does not have the same number of black and white stones.

== Links ==

'''Program:'''

* David Wu's original KataGo: <https://github.com/lightvector/KataGo>
* HZY's KataHex: <https://github.com/hzyhhzy/KataGo/tree/Hex2022>
* Selinger's moficiations: <https://github.com/selinger/katahex>

'''Pre-trained network models:'''

* HZY's pre-trained 19x19 network: [https://drive.google.com/file/d/1xMvP_75xgo0271nQbmlAJ40rvpKiFTgP/view katahex_model_20220618.bin.gz]

'''GUIs:'''

* HZY's KataHex works with this modified LizzieYzy GUI: [https://drive.google.com/file/d/1qbTTmPFiUkM_346DeKS1E9gJR-roNH63/view KataHex_LizzieYZY_20220313.zip].
* Selinger's KataHex works with HexGUI: <https://github.com/selinger/hexgui>

'''Analysis:'''

* HZY's analysis of win-rates for opening moves: [https://zhuanlan.zhihu.com/p/476464087 Chinese] [https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en English]
* [[Strategic advice from KataHex]]

[[Category:Computer Hex]]
[[category:Hex playing program]]

@@ Line 9: / Line 9: @@
 == Pre-trained networks ==
-HZY initially trained the neural network model on two NVIDIA GeForce RTX 2080 Ti GPUs for about 20 days on 13x13. They then up-trained the network on 19x19 for one day, and on 27x27 for an additional day. [https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en According to HZY], the up-trained 19x19 network is relatively reliable, but the 27x27-network is not.
+HZY initially trained the neural network model (20220618) on two NVIDIA GeForce RTX 2080 Ti GPUs for about 20 days on 13x13. They then up-trained the network on 19x19 for one day, and on 27x27 for an additional day. [https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en According to HZY], the up-trained 19x19 network is relatively reliable, but the 27x27-network is not.
-The newest model, 20240812, is trained on two RTX 4090 GPUs for 2 months on 15x15, 15 days on 19x19, and 3 days on 27x27.
+A newer model, 20240812, is trained on two RTX 4090 GPUs for 2 months on 15x15, 15 days on 19x19, and 3 days on 27x27. This model uses a b18c384nbt architecture (18 block, 384 channel, [https://github.com/lightvector/KataGo/blob/master/docs/KataGoMethods.md#nested-bottleneck-residual-nets nested bottleneck residual net]).
 Running KataHex requires both a neural net model (the "weights"), and an engine to load the weights. Each neural net is able to play Hex not only at the size it was trained on, but also smaller and larger sizes, as long as the engine is compiled to support the size. (However, the net will not play particularly well on sizes larger than it was trained on.)
@@ Line 25: / Line 27: @@
 '''Program:'''
-* David Wu's original KataGo: <https://github.com/lightvector/KataGo>
+* David Wu's original KataGo: https://github.com/lightvector/KataGo
-* HZY's KataHex: <https://github.com/hzyhhzy/KataGo/tree/Hex2024>
+* HZY's KataHex: https://github.com/hzyhhzy/KataGo/tree/Hex2024
-* Selinger's modifications: <https://github.com/selinger/katahex/tree/Hex2024>
+* Selinger's modifications: https://github.com/selinger/katahex/tree/Hex2024
 '''Pre-trained network models:'''
-* 20250131 (current strongest net): [https://github.com/hzyhhzy/KataGomo/releases/download/Hex_20250131/hex3_27x_b28.bin.gz hex3_27x_b28.bin.gz]
+* 20250131 (latest): https://github.com/hzyhhzy/KataGomo/releases/Hex_20250131
-* 20250131 (supports move limit mode, but slightly weaker): [https://github.com/hzyhhzy/KataGomo/releases/download/Hex_20250131/hex3_mm19x_b28.bin.gz hex3_mm19x_b28.bin.gz]
+** (weights only): [https://github.com/hzyhhzy/KataGomo/releases/download/Hex_20250131/hex3_27x_b28.bin.gz hex3_27x_b28.bin.gz]
-* 20240812: [https://drive.usercontent.google.com/download?id=1YeqRvAYs7YjtPh0xBbDnHxeo2xrLEOdX hex27x3.bin.gz]
+** (supports move limit mode, but slightly weaker): [https://github.com/hzyhhzy/KataGomo/releases/download/Hex_20250131/hex3_mm19x_b28.bin.gz hex3_mm19x_b28.bin.gz]
 * 20220618: [https://drive.google.com/file/d/1xMvP_75xgo0271nQbmlAJ40rvpKiFTgP/view katahex_model_20220618.bin.gz]
 '''GUIs:'''
-* The 20240812 net is bundled with a modified LizzieYzy GUI: <https://github.com/hzyhhzy/KataGo/releases/tag/Hex_20240812>
+* The 20240812 net is bundled with a modified LizzieYzy GUI: https://github.com/hzyhhzy/KataGo/releases/tag/Hex_20240812
-* An older net with the modified LizzieYzy GUI: [https://drive.google.com/file/d/1qbTTmPFiUkM_346DeKS1E9gJR-roNH63/view KataHex_LizzieYZY_20220313.zip].
+* An older net with the modified LizzieYzy GUI: [https://drive.google.com/file/d/1qbTTmPFiUkM_346DeKS1E9gJR-roNH63/view KataHex_LizzieYZY_20220313.zip]
-* Selinger's KataHex works with HexGUI: <https://github.com/selinger/hexgui>
+* Selinger's KataHex works with HexGUI: https://github.com/selinger/hexgui
 == See also ==

@@ Line 31: / Line 31: @@
 '''Pre-trained network models:'''
-* 20240812 (current strongest net): [https://drive.usercontent.google.com/download?id=1YeqRvAYs7YjtPh0xBbDnHxeo2xrLEOdX hex27x3.bin.gz]
+* 20250131 (current strongest net): [https://github.com/hzyhhzy/KataGomo/releases/download/Hex_20250131/hex3_27x_b28.bin.gz hex3_27x_b28.bin.gz]
 * 20220618: [https://drive.google.com/file/d/1xMvP_75xgo0271nQbmlAJ40rvpKiFTgP/view katahex_model_20220618.bin.gz]

@@ Line 26: / Line 26: @@
 * David Wu's original KataGo: <https://github.com/lightvector/KataGo>
-* HZY's KataHex: <https://github.com/hzyhhzy/KataGo/tree/Hex2022>
+* HZY's KataHex: <https://github.com/hzyhhzy/KataGo/tree/Hex2024>
-* Selinger's modifications: <https://github.com/selinger/katahex>
+* Selinger's modifications: <https://github.com/selinger/katahex/tree/Hex2024>
 '''Pre-trained network models:'''

@@ Line 31: / Line 31: @@
 '''Pre-trained network models:'''
-* 20240812 (current strongest net): <https://github.com/hzyhhzy/KataGo/releases/tag/Hex_20240812>
+* 20240812 (current strongest net): [https://drive.usercontent.google.com/download?id=1YeqRvAYs7YjtPh0xBbDnHxeo2xrLEOdX hex27x3.bin.gz]
 * 20220618: [https://drive.google.com/file/d/1xMvP_75xgo0271nQbmlAJ40rvpKiFTgP/view katahex_model_20220618.bin.gz]

@@ Line 9: / Line 9: @@
 == Pre-trained networks ==
-HZY initially trained the neural network model on two NVIDIA GeForce RTX 2080 Ti GPUs for about 20 days. They then up-trained the network on 19x19 for one day, and on 27x27 for an additional day. [https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en According to HZY], the up-trained 19x19 network is relatively reliable, but the 27x27-network is not.
+HZY initially trained the neural network model on two NVIDIA GeForce RTX 2080 Ti GPUs for about 20 days on 13x13. They then up-trained the network on 19x19 for one day, and on 27x27 for an additional day. [https://zhuanlan-zhihu-com.translate.goog/p/476464087?_x_tr_sl=auto&_x_tr_tl=en According to HZY], the up-trained 19x19 network is relatively reliable, but the 27x27-network is not.
 Each pre-trained neural network is able to play Hex at the size it was trained on, or any smaller size. Thus, the 19x19 network model can also be used to play 11x11 or 13x13.

@@ Line 40: / Line 40: @@
 * Selinger's KataHex works with HexGUI: <https://github.com/selinger/hexgui>
-'''Analysis:'''
+== See also ==
 * [[Strategic advice from KataHex]]

@@ Line 19: / Line 19: @@
 == Limitations ==
-Since KataHex was only trained on self-play, it does not always do well when asked to play from an arbitrary starting position. It seems to be particularly thrown off if the starting position does not have the same number of black and white stones.
+Since KataHex was only trained on self-play, it does not always do well when asked to play from an arbitrary starting position. Older nets are particularly thrown off if the starting position does not have the same number of black and white stones, but the issue seems to have been mitigated with the 20240812 net.
 == Links ==
@@ Line 27: / Line 27: @@
 * David Wu's original KataGo: <https://github.com/lightvector/KataGo>
 * HZY's KataHex: <https://github.com/hzyhhzy/KataGo/tree/Hex2022>
-* Selinger's moficiations: <https://github.com/selinger/katahex>
+* Selinger's modifications: <https://github.com/selinger/katahex>
 '''Pre-trained network models:'''
@@ Line 37: / Line 37: @@
 * The 20240812 net is bundled with a modified LizzieYzy GUI: <https://github.com/hzyhhzy/KataGo/releases/tag/Hex_20240812>
-* An older net with the LizzieYzy GUI: [https://drive.google.com/file/d/1qbTTmPFiUkM_346DeKS1E9gJR-roNH63/view KataHex_LizzieYZY_20220313.zip].
+* An older net with the modified LizzieYzy GUI: [https://drive.google.com/file/d/1qbTTmPFiUkM_346DeKS1E9gJR-roNH63/view KataHex_LizzieYZY_20220313.zip].
 * Selinger's KataHex works with HexGUI: <https://github.com/selinger/hexgui>