サッカーにおけるポジショニングの修正を行う遺伝的アルゴリズムモデルの提案

サッカーのポジショニングは，少しの違いで失点のリスクが大きく変わるため解釈が難しい．そこで本研究では，ポジショニングを数値的に評価する手法（Pitch Risk）を導入した，ポジショニング修正モデルを提案する．アマチュアの選手の試合後の振り返りでの利用などを想定している．Pitch Riskは，ピッチの支配状況を示すPitch Control，ゴールの確率を示すxG，アシストの確率を示すSxAを用いて，確率計算を基に算出した．サッカー経験者7人が局面において失点のリスクが高いと回答したエリアとのJS-divergenceは，Pitch Controlの118.18に対して，Pitch Riskは0.14であることから，Pitch Riskは失点のリスクをより表現できている．ポジショニング修正モデルにはGAを用いた．守備側11選手の距離と方向の計22変数を遺伝子とし，評価関数にはPitch Riskを用いた．GAで生成した個体数分ランダム生成した結果と比較すると，約3%の数値的改善が生じた．また，11選手分の学習を同時に進めたことで選手が連動した改善が見られた．モデル改善による精度向上により，ポジショニングの解釈を幅広い人々に提供できると考える．

The positioning of soccer is difficult to interpret because a little difference can significantly change the risk of conceding a goal. Therefore, we introduce a method for quantifying the risk of conceding a goal (Pitch Risk) and propose a model to output the modification of positioning in a phase based on the method. The model is intended for use by amateur athletes in their post-game review. Pitch Risk is calculated based on a fundamental rule of probability using Pitch Control, which indicates pitch dominance, xG, which indicates the probability of a goal, and SxA, which indicates the probability of an assist. Pitch Risk is a better representation of the risk of conceding a goal since the JS-divergence with the areas where the seven experienced soccer players reported a higher risk of conceding a goal in the phase is 0.14 for Pitch Risk compared to 118.18 for Pitch Control. We use GA for the Positioning Modification model. A total of 22 variables of distance and direction for the 11 defending players were used as chromosomes, and Pitch Risk was used as the evaluation function. Improvement of about 3% occurred when compared to the best of randomly generated. In addition, there was a player-linked improvement due to the simultaneous learning of 11 players. We believe that the improvement of accuracy for the model provides positioning interpretations to a wider number of people.