Evaluating Go game records for prediction of player attributes

Abstract

Abstract-We propose a way of extracting and aggregating permove evaluations from sets of Go game records. The evaluations capture different aspects of the games such as played patterns or statistic of sente/gote sequences. Using machine learning algorithms, the evaluations can be utilized to predict different relevant target variables. We apply this methodology to predict the strength and playing style of the player (e.g. territoriality or aggressivity) with good accuracy. We propose a number of possible applications including aiding in Go study, seeding realwork ranks of internet players or tuning of Go-playing programs.

Evolving Non-Linear Stacking Ensembles for Prediction of Go Player Attributes

“…The prediction of player attributes as demonstrated in this work has been (together with the feature extraction presented in [2]) combined in an online web application 1 , which evaluates games submitted by players and predicts their playing strength and style. The predicted strength is then used to recommend relevant literature and the playing style is utilized by recommending relevant professional players to review.…”

“…Prediction of Go player attributes has until recently been limited to pre-defined questionnaires and simple methods [14], [15]. Universal approach to the problem has been introduced by our previous work [2] and [3].…”

“…Dataset: We have collected a large sample of games from the public archives of the Kiseido Go server [23], the sample consists of over 100 000 records of games. The records were divided by player's strength and preprocessed as is detailed in [2]. Here, it is enough to note that the dataset consists of 120 independent pairs (x, y) for each of the 26 ranks, where x is the feature vector described in [2] (dimension of x was 1040), and y is the target variable, which is one number describing the rank .…”

“…For each of 24 professional players, we obtained 12 pairs (x, y), where x is the feature vector obtained from the games as described in [2] (dimension of x was 640), and y is one of the 4 styles-basically, we view the problem as 4 different regression problems which share the same feature vectors.…”

“…We focus on analyzing existing game records with the aim of helping humans to play and understand the game better instead. In our previous work [2], we have presented a way to extract information rich features from sets of Go players' games.…”