A graph-integrated reinforcement learning framework with graph neural networks for tactical decision modeling in professional football.

The increasing availability of high-frequency player tracking and event data has enabled more detailed analysis of football tactics, yet many existing approaches struggle to connect learned decisions with realistic execution constraints and interpretable tactical reasoning. This study proposes a graph-integrated reinforcement learning framework with graph neural networks (GIRL-GNN) for offline tactical decision modeling in professional football. Each decision moment is represented as a heterogeneous graph constructed from synchronized tracking and event data, encoding player interactions, ball movement, spatial context, and match state. A hierarchical policy separates high-level tactical intent from low-level executable control, while an explicit feasibility module evaluates candidate actions with respect to timing, visibility, defensive pressure, and rule constraints. Learning combines calibrated predictions of value gain, defensive risk, and execution feasibility under a risk-sensitive objective. The framework is evaluated on seven professional matches using a leave-one-match-out protocol. Results indicate consistent improvements over sequence-based and rule-based baselines across multiple tactical dimensions, including offensive progression, defensive suppression, feasibility, and robustness under input noise within the evaluated dataset. Calibration analysis indicates reliable alignment between predicted and observed outcomes, and qualitative analyses show that the learned policies reflect recognizable tactical structures consistent with professional performance indicators. These findings suggest that feasibility-aware, graph-based reinforcement learning provides a promising direction for data-driven tactical decision analysis within the scope of the evaluated dataset.