Relational multi agent tactical learning for competitive football

Abstract

Tactical decision-making of competitive football depends on the rapidly changing spatial relationship, the transformation of ball rights and the fight against oppression. The existing multi-agent reinforcement learning methods often compress the team state into fixed observation or pooled representation, and the node-level collaborative information is weakened in strategy learning; When the opponent’s style changes, the strategy execution is also prone to degradation. Aiming at the above problems, a relational multi-agent tactical learning method for competitive football is proposed. Player interaction is modeled as a dynamic graph structure, relationship coding is used to describe local cooperation, pressing pressure and passing channels, and time series modeling is used to capture tactical phase transition. Shared cooperative mechanism, self-game training and meta-adaptive strategy are brought into the same playback learning process to improve the consistency of strategies, confrontation stability and cross-scene adaptability. Based on StatsBomb and SoccerNet datasets, experiments are carried out in the tasks of attack, defense, counterattack and positioning. The results show that, in the evaluated open dataset and replay tactical environment, the highest tactical execution score of the proposed method reaches 82.6 in complex tactical tasks, and it maintains a relatively stable strategic adaptation performance in cross-scene migration tasks. The research results show that retaining the cooperative information of player relationship is helpful to improve the quality of dynamic multi-agent decision-making in replay football tactical learning.