BISAK COM059000 Computer Engineering BISAK COM012000 Computer Graphics BISAK COM014000 Computer Science BISAK COM072000 Computer Simulation BISAK COM017000 Cybernetics BISAK COM062000 Data Modeling & Design BISAK COM018000 Data Processing BISAK COM021030 Databases / Data Mining BISAK COM071000 Digital Media / Video & Animation BISAK COM023000 Educational Software BISAK COM025000 Expert Systems BISAK COM074000 Hardware / Mobile Devices BISAK COM032000 Information Technology BISAK COM077000 Mathematical & Statistical Software BISAK BUS021000 Econometrics BISAK GAM001000 Board Games BISAK GAM016000 Fantasy Sports BISAK GAM004050 Gambling / Sports BISAK EDU039000 Computers & Technology BISAK EDU036000 Organizations & Institutions BISAK EDU037000 Research BISAK EDU027000 Statistics BISAK MED084000 Sports Medicine BISAK MED078000 Public Health BISAK MED076000 Preventive Medicine BISAK HEA046000 Children's Health BISAK HEA006000 Diet & Nutrition / Diets BISAK HEA034000 Diet & Nutrition / Food Content Guides BISAK HEA002000 Aerobics BISAK HEA017000 Diet & Nutrition / Nutrition BISAK HEA007000 Exercise BISAK HEA010000 Healthy Living BISAK HEA038000 Work-Related Health BISAK MAT011000 Game Theory BISAK MAT029030 Probability & Statistics / Regression Analysis BISAK MAT025000 Recreations & Games BISAK SPO003000 Baseball / General BISAK SPO068000 Business Aspects BISAK SPO077000 Children’s & Youth Sports BISAK SPO061020 Coaching / Football BISAK SPO061030 Coaching / Soccer BISAK SPO011000 Cycling BISAK SPO063000 Equipment & Supplies BISAK SPO073000 Field Hockey BISAK SPO015000 Football BISAK SPO016000 Golf BISAK SPO075000 Health & Safety BISAK SPO024000 Juggling BISAK SPO040000 Soccer BISAK SPO066000 Sociology of Sports BISAK SPO047000 Training BISAK SPO051000 Water Sports BISAK PSY012000 Education & Training BISAK PSY042000 Assessment, Testing & Measurement
Abstract and keywords
Abstract (English):
The article describes testing a weakness of the go-playing AI KataGo (9d) detected by K. Pelrine et al. Building groups following the developed 4-level algorithm reveals the incapability of AI to understand intentional sacrifice of larger groups and to prioritize when discovering several similar sacrifices on the board. The value of the developed strategy lies in human victory over a potentially unbeatable program even regarding the fact that human is playing on the level of 1q (2000 Elo).

Keywords:
go game, baduk, weiqi, KataGo, neural networks
Text
Text (PDF): Read Download
References

1. AlphaGo by Google DeepMind. URL: https://www.deepmind.com/research/highlighted-research/alphago (accessed 20.09.2023).

2. KataGo on GitHub. URL: https://github.com/lightvector/KataGo/releases?q=v1.12.4&expanded=true (accessed 20.09.2023).

3. Leela official website. URL: https://www.sjeng.org/leela.html (accessed 20.09.2023).

4. Lauder E. Tencent's Made a Go-Playing AI to Rival Google's AlphaGo. URL: https://aibusiness.com/companies/tencent-s-made-a-go-playing-ai-to-rival-google-s-alphago (accessed 20.09.2023).

5. Wang T.T., Gleave A., Tseng T., Pelrine K., Belrose N. et al. Adversarial policies beat superhuman Go AIs // arXiv:2211.00241 [cs.LG]. URL: https://goattack.far.ai/adversarial-policy-katago (accessed 20.09.2023). DOI:https://doi.org/10.48550/arXiv.2211.00241

6. Financial Times. Man beats machine at Go in human victory over AI. URL: https://www.ft.com/content/175e5314-a7f7-4741-a786-273219f433a1 (accessed 27.11.2023).

Login or Create
* Forgot password?