These stats were calculated using the Go Performance Quality (GPQ) algorithm. Joseph Craig, AKA sadakatsu, created this algorithm to generate a single value to represent how well a player played. Other metrics require multiple values to provide context: mean point loss needs at least the standard deviation, accuracy needs best move percentage and match percentage, and tracking moves' win rates or point losses need all the other moves' values to tell one anything. What are the numbers? The Quality Score (biggest number on the top) uses a player's whole set of point loss values to generate one score that can indicate three values simultaneously: how many mistakes the player made during the gamehow many points the player lost during the gameroughly how strongly the player performed These have been tuned to set 0— as a theoretical worst player— and 100— as KataGo at peak performance. As you will see, the games our candidates graciously shared with us for this project show that they have several stats in the 70's and higher, showing that they are near our other NAGF players in skill. Next is the Average Rating, which is not their AGA or CGA rating, but rather the algorithms suspected rating averaged across all of their games. Then in the bottom we have a myriad of abbreviations for the sake or spacing. Starting on the left side: OVR: Overall, uses every moveBLK: Black, uses every move as BlackWHT: White, uses every move as WhiteOPN: Opening, uses first 50 moves of each game (1-50)MID: Middle game, uses the next 100 moves of each game (51-150)END: Endgame, uses every move from the rest of the game (151+) On the bottom right-hand side are the players calculated GoStyle's, which is a wonderful tool developed by Josef Moudřík and Petr Baudiš and can be found at http://gostyle.j2m.cz/. To be clear, neither the left style nor the right style is better, it is merely a continuum between the two. The styles are Moyo: Influence style playerTRTY: Territorial style playerCLSC: Classical move styleNOVL: Novel or more modern move styleCALM: Point focus styleFGHT: Fighting styleLGHT: Light positions, hard to killTHCK: Thick and solid, but less safe How was this made? When asked how the algorithm works, sadakatsu will start to rattle off a chain of applied statistical techniques until the listener gets bored and walks away. He offered to use his GPQ algorithm to generate Quality Scores to use as stats for the tournament's candidates and several other reference players. He generated these stats to discover how well they play in different areas of the game using up to 100 of the players' most recent games on Fox. Some players were only able/willing to share a small handful of games, and it should be noted that the statistics for those players are less likely to be accurate. Using KataGo's recently added human profile capability, sadakatsu also generated stone level performance ratings for the players' skill levels. Incredibly, most players performed very closely in their online Average Performance Ratings to their official national organization's ratings. Can I analyze my games? This tool's core is currently available for free use at https://github.com/sadakatsu/go-performance-quality. You can calculate these stats and more for your games. It also includes a new GoStudy tool. While still prototypical, it has proven incredibly effective in focusing a player's review on the game on moves that matter at that player's level of play. It separates moves that make sense at that player's performance level from the ones the AI recommends when the AI wants something too tricky. It also filters down the moves to review to a manageable number to help students spend their time well. Joseph (sadakatsu) and Jean-François Gagnier are currently working to deploy this tool as a website that will be called The Go Database. They are working to get this site live for an alpha before the end of the year. If you want to participate in a testing round or want status updates on the project, please e-mail jacraig@sadakatsu.com to let us know how you would like to participate. Credits A huge thanks to Joseph (sadakatsu) for the many hours put into coding and running these analyses and to Jean-François Gagnier for his ideas and insights which were key in moving this project from a nice idea through a confusing mess of numbers to a form that we think our audience can enjoy. Time and life interfered from using all Gagnier's contributions for this project, but they help turn The Go Database into something incredible. Thanks again to Josef Moudřík and Petr Baudiš for their wonderful GoStyle tool. Thanks to 'curlywalrus' for his help getting the games off of Fox. A huge thanks to all the players who sent us their games and photos to use for this. And thanks to Nick Prince of the NAGF connecting us with those players when others in charge ignored my requests. And in case you're curious I did the graphic design. *pats self on back* This year for the NAGF Pro Qualifications I wanted to create some awesome player cards with real statistics about their abilities and play styles. I'd found a fun chess example on reddit and thanks to the power of Twitch chat and friendship I was able to connect with a terrific software engineer and a statistics loving quantitative analyst.
