Quantitative Comparative Analysis of Traditional Joseki and AI-Recommended Moves

Kim Jaeyun /June 20, 2026

Abstract
AlphaGo’s landmark victory over Lee Sedol in March 2016 triggered an unprecedented paradigm shift in the game of Go,prompting widespread revaluation of joseki sequences—locally optimal opening patterns refined over centuries of human tradition.Despite this upheaval,systematic quantitative research into precisely how inefficient traditional joseki are—measured in concrete point(目)differentials—remains scarce in the academic literature. Most existing discourse has operated at the level of qualitative judgment (“this move is good/bad”)without rigorously measuring the numerical stakes.

This study addresses that gap by extracting approximately 70 key moves from 20 fuseki patterns widely used in the pre-AlphaGo era and quantifying the efficiency difference between traditional sequences and AI-recommended moves using KataGo’s Expected Score metric.A central contribution is the independent design and development of Joseki Analyzer—a purpose-built program integrating a FastAPI backend with the KataGo engine—enabling automated,large-scale,reproducible analysis under standardized conditions (1,000 visits,Chinese rule set,komi 7.5).The core metric Δ Score is defined as the Score Lead of the AI’s top recommended move minus the Score Lead of the traditional move at the same position;a negative value indicates that the traditional move is less efficient by the corresponding number of points.

Results show an overall mean Δ Score of approximately −0.28 points across 20 patterns,indicating that traditional moves incur an expected-score loss of roughly this magnitude per move relative to AI recommendations. The largest divergences occur in the Komoku Approach–Pincer Response (Ⅱ)(−1 .20 pts),Komoku Approach Aggressive Response (−0 .68),Hoshi One Space Pincer–3-3 Invasion (−0 .60),and Komoku Corner Enclosure Fuseki and Komoku Approach–High Extension (both −0.53).The single largest move-level loss is −2 .59 points.Conversely,four patterns achieve Δ Score = 0—Hoshi Approach–Knight’s -Move Response, Komoku Enclosure–Development Variation, Hoshi Approach–Contact-Play Joseki,and Komoku Approach–Contact Play (Ⅱ)—indicating perfect alignment with AI evaluation.

A consistent typological finding emerges:corner-enclosure and extension patterns show the largest divergence from AI,while contact-play (붙임수) patterns show the smallest.Across all patterns,KataGo systematically prioritizes claiming empty corners over reinforcing one’s own established positions—a finding that runs counter to a core axiom of classical Go strategy.This study represents the first systematic,tool-assisted effort to quantify the inefficiency of traditional fuseki joseki in point-based terms,offering both empirical findings and a replicable methodological framework for evaluating classical Go theory against modern AI computation.

Keywords: Go,joseki,fuseki,KataGo,Expected Score,quantitative analysis,AlphaGo,AI efficiency,Δ Score,Joseki Analyzer