AI Learns to Judge Every Soccer Pass by Reading the Whole Pitch

Every soccer match produces thousands of small decisions: pass or dribble, shoot or hold, press or drop. Judging whether those decisions were good has long been one of the hardest problems in sports analytics, because the value of any single action depends on everything around it — where teammates and opponents stand, how the possession unfolded, and what the defending team is likely to do next. A new study by Maryam Tamimi and Thomas Tran of the University of Ottawa, published in the International Journal of Data Science and Analytics, tackles this problem with a machine learning framework that treats the two teams as separate strategic agents and learns to value actions directly from logged match data.

The researchers built their system around offline reinforcement learning, a branch of AI in which an algorithm learns decision-making policies from a fixed dataset of past behavior rather than by trial and error in a live environment. This matters for soccer because teams cannot be used as experimental sandboxes: the only data available are matches that have already been played. The framework draws on publicly available StatsBomb event data and StatsBomb 360 freeze-frame annotations, which record not just what happened on the ball but the positions of all players at key moments, giving the model a full spatial snapshot of the game.

To make sense of that spatial information, the authors encode player, ball, and goal configurations into image-like tensors and process them with a ResNet-based spatial encoder, a deep neural architecture originally developed for image recognition. Each offensive and defensive state is compressed into a 64-dimensional latent vector that captures the tactical geometry of the moment. The temporal dimension is handled by Transformer-based encoders, the attention-driven architecture that underlies modern language models, allowing the system to model possession-level action sequences and understand how earlier events shape the value of later ones.

The central innovation, however, is the dual-agent structure. Instead of modeling soccer as a single-agent problem in which opponents are just part of the environment, the framework represents the possessing team and the opponent as separate agents, each with its own policy, action space, and Q-value function — the learned estimate of the long-term value of taking a given action in a given state. This opponent-aware design means that when the model evaluates a pass or a dribble, it explicitly accounts for how the defending team’s choices influence the outcome, a closer approximation of the adversarial nature of real match play.

Working with offline data introduces a well-known hazard: reinforcement learning agents can learn to overestimate the value of actions they have rarely seen, producing confident but unreliable recommendations. Tamimi and Tran address this with several stabilization techniques. Behavior-support masking restricts learning to actions that are actually supported by the logged behavior policy, preventing the model from assigning values to moves that never realistically occur in the data. KL regularization keeps the learned policy close to the observed behavior distribution, and phase-aware optimization tailors the learning process to distinct tactical phases of play such as build-up, established possession, and attack.

The preprocessing pipeline reflects the same attention to detail. Raw event and freeze-frame data are sorted by match, period, and timestamp, and grouped into possessions. Possessions with fewer than ten events are discarded, while those longer than 300 events are truncated to their most recent 300. Ball locations are assigned to one of 20 pitch zones using nearest-centroid assignment, and those same zones define the tactical phases. Rewards are normalized at the possession level and clipped, producing standardized training episodes in which each timestep contains offensive and defensive state embeddings, actions, rewards, behavior-policy probabilities, and ball-zone labels.

The empirical results show that the framework learns what the authors describe as conservative but value-directed policy adjustments. Rather than proposing radical departures from how teams actually play, the model nudges its estimated policies toward higher-value actions while remaining anchored to the behavior-supported distribution. Notably, the largest improvement opportunities concentrate in attacking phases, suggesting that the final third of the pitch — where decisions are most consequential and most contested — is where better action valuation offers the greatest analytical payoff.

Training statistics support the picture of a stable, disciplined learner. Across four cross-validation folds, the validation EPO — the estimated possession outcome — rises from 0.1914 to 0.1944 over five epochs while the validation log-loss remains steady. That combination indicates the policy improves its estimated possession value without drifting away from the behavior-supported action distribution. Evaluation relies on weighted importance sampling, Q-based diagnostics, KL divergence from behavior, effective sample size, and bootstrap confidence intervals, a battery of techniques designed to make offline policy assessment as trustworthy as possible.

The work sits within a rapidly growing research landscape. Earlier approaches such as VAEP, which values player actions in soccer by modeling how actions change the probability of scoring, and multi-agent deep reinforcement learning frameworks for valuing on- and off-ball movement laid important groundwork. Recent studies have applied Transformers to soccer event prediction and marked spatio-temporal point processes to match analysis. What distinguishes the new framework is its combination of spatiotemporal representation learning with opponent-aware dual-agent offline RL, explicitly built for behavior-constrained policy valuation from logged data rather than simulated play.

For clubs, analysts, and researchers, the implications are practical. A stable, opponent-aware action valuation model could support scouting, tactical review, and player evaluation from data that many organizations already collect, without requiring access to proprietary tracking feeds. The authors note that the code will be made available upon request, and the underlying datasets are freely available through StatsBomb Open Data, lowering the barrier for replication. As sports analytics matures, frameworks like this one point toward a future in which the quality of every touch on the ball can be measured against the full spatial and strategic context in which it occurred — and in which the defending team’s intelligence is no longer an afterthought but a first-class citizen of the model.

Subject of Research: Offline reinforcement learning for soccer action valuation with spatial-temporal modeling

Article Title: Support-aware dual-agent reinforcement learning for soccer possession and action valuation with spatial–temporal modeling

Article References: Tamimi, M., & Tran, T. (2026). Support-aware dual-agent reinforcement learning for soccer possession and action valuation with spatial–temporal modeling. International Journal of Data Science and Analytics, 22(1), Article 343. https://doi.org/10.1007/s41060-026-01328-x

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01328-x

Keywords: soccer analytics, reinforcement learning, offline reinforcement learning, action valuation, spatial-temporal modeling, Transformer, ResNet, StatsBomb, dual-agent learning, sports data science, possession value, machine learning