Kata Go

    Keywords: Software

KataGo has been the state-of-the-art generally available[1] Go-playing AI system since 2020.

Table of contents

Announcement

From its [ext] reference announcement on github:

Research and experimentation with self-play training in Go. Contains a working implementation of AlphaZero-like training with a lot of modifications and enhancements. Due to these enhancements, early training is immensely faster than in other zero-style bots – with a only few strong GPUs for a few days, even a single person should be able to train to mid or even high amateur dan strength on the full 19x19 board on consumer hardware. Also contains a GTP engine and pre-trained neural nets competitive with other top open-source Go engines. KataGo is also capable of estimating score and territory, and due to this right-out-of-the-box plays handicap games somewhat better than many other zero trained bots, without any special hacks.

Related Software and Manuals

  • [ext] Lizzie: KataGo is one of the two engines usable in the analysis program Lizzie, which now (from version 0.7.1 on) displays a graph of and shifts in KataGo’s score estimates. The default engine remains Leela Zero, which, however, does not provide a score estimate.
  • [ext] LizzieYzy is a graphical interface modified from Lizzie,allow loading various engines like: Katago、LeelaZero、Leela、ZenGTP、SAI、Pachi or other GTP engines.
  • [ext] KaTrain is a program that allows one to use KataGo for training: one can analyse games, play against various artificially weakened versions of KataGo, optionally getting bad moves pointed out as you play. The weakened bots are also available as opponents on OGS under the names katrain-*.
  • [ext] LizGoban is an analysis tool which uses LeelaZero and KataGo. UI is built using Electron.
  • [ext] Ogatak is a relatively simple KataGo-specific GUI. UI is built using Electron.
  • AI Sensei is a website that offers free !Katago game reviews and analysis, created by Benjamin Teuber.
  • [ext] BadukAI is the first native implementation of playing with modern Go Ais on android.
  • [ext] AI Computer is a manual by Robert Jasiek for installation of Katago, Nvidia libraries and GUIs.

Adversarial attacks

Although KataGo is extremely strong, it proved possible to create and train a specialised “adversarial” bot which, although weaker than 20 kyu, took advantage of access to KataGo’s network (without the weights) to defeat KataGo most of the time.

Premature passing

Many people initially supposed that this was due to some intentional exploitation of the game scoring mechanism, as KataGo passed without capturing all opposing stones, which is a mistake under Tromp-Taylor Rules. It turns out, however, that the bot genuinely found and exploited a weakness in KataGo, which is trained to play against good moves unlike those the bot plays, and is therefore capable of serious misjudgements in other situations.

[ext] KataGo v1.12.0 repaired this weakness.

sample game
Premature passing


Encircling the encircler

A second weakness discovered and exploited by the bot is that KataGo does not always recognise that it must defend one of its groups, when that group has surrounded an enemy group, but is itself surrounded.

As of February 2023 KataGo was being trained on such cyclic positions.

sample game
The encircler encircled


Links

Notes

[1] it is conceivable that there are stronger programs not publicly available, though I, PJT, suspect not, or that they are based on KataGo.


Kata Go last edited by 213.125.2.98 on October 2, 2023 - 16:10
RecentChanges · StartingPoints · About
Edit page ·Search · Related · Page info · Latest diff
[Welcome to Sensei's Library!]
RecentChanges
StartingPoints
About
RandomPage
Search position
Page history
Latest page diff
Partner sites:
Go Teaching Ladder
Goproblems.com
Login / Prefs
Tools
Sensei's Library