🔢 mathematics

Constrained Multi-Agent Optimal Control for Football Tactical Decision Support: Analysis, Solver Verification, and Reproducible Scenario Experiments

This paper presents a constrained multi-agent optimal control framework for football tactical decision-making that establishes theoretical optimality conditions, implements a numerical solver, and validates the approach through reproducible synthetic experiments demonstrating significant improvements in strategic performance over baseline strategies.

Original authors: Abbes Kadyri, Abdelfettah Belkahla, Otmane Bencheihb

Published 2026-09-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abbes Kadyri, Abdelfettah Belkahla, Otmane Bencheihb

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Football is often described as a game of space and time, where the most valuable asset is not just the ball, but the empty ground surrounding it. For decades, coaches and analysts have watched replays to understand how teams move together, trying to spot patterns in how a group of eleven players shifts from a tight defensive block to an attacking formation. Modern technology has made this observation even sharper, allowing us to track the exact position of every player on the pitch at high speed. However, knowing where players are and how they moved in the past is different from knowing what they should do next. The big question that remains is how to calculate the perfect path for a whole team to move from one shape to another, while respecting the physical limits of the players and the rules of the field. This is the challenge of tactical decision support: turning the abstract idea of a "good move" into a concrete, executable plan.

A team of researchers at Chouaib Doukkali University in Morocco has tackled this problem by treating football tactics as a complex math puzzle. They built a computer model that simulates a full match scenario, not by guessing what might happen, but by calculating the single best way for a team to rearrange itself. In their study, they did not rely on real-world match footage or historical data. Instead, they created a synthetic, or entirely made-up, scenario involving eleven players on a standard field measuring 105 by 68 meters. They defined a set of goals that a coach might want to achieve, such as controlling more of the field, keeping the defensive line tight, pressing the opponent aggressively, and conserving energy. They also set strict physical rules, ensuring that no player in the simulation could run faster than a realistic maximum speed of 7.0 meters per second.

The researchers asked their computer to find the optimal path for every player to travel from a starting position to a target position over a period of twelve seconds. To do this, they used a method that constantly checks the trade-offs between different objectives. For instance, if a team tries to press the opponent harder, the players must run faster, which uses more energy. If they stay too far apart, they lose defensive stability. The computer had to balance all these competing needs simultaneously. The result was a set of precise instructions for how each player should move at every fraction of a second to achieve the best overall outcome.

When the researchers compared the computer's optimized plan against two simpler alternatives, the difference was clear. The first alternative was a static baseline, where the team simply stayed in their starting formation without moving. The second was a linear transition, where players moved in a straight line at a constant speed toward their target, like a simple animation. The computer's optimized solution, which involved complex, curved paths where players accelerated and slowed down at specific moments, performed significantly better. It reduced the overall "cost" of the maneuver by 11.58 percent compared to the simple straight-line movement and by 38.26 percent compared to staying still. This improvement came from the model's ability to position players so they could cover more ground and apply more pressure, even though it required them to run slightly harder.

The study also tested how the team would behave if the coach changed their priorities. In one experiment, the researchers told the computer to focus almost entirely on pressing the opponent. The resulting plan showed the players moving more aggressively to close down space, accepting a higher risk of leaving gaps in defense. In another experiment, they prioritized defensive stability. The computer responded by having the players move more slowly and stay closer together, reducing their energy expenditure but allowing the opponent more space. These tests proved that the system was not just following a rigid script; it was genuinely solving a balancing act, shifting its behavior based on the specific instructions given.

It is important to understand the limits of this work. The researchers were very clear that this was a mathematical verification, not a prediction of real-life football. The players in the simulation were not real humans with fatigue, injuries, or individual quirks; they were idealized points moving according to strict rules. The opponent was also fixed and predictable, not a thinking adversary trying to outsmart the team. The study did not claim to have solved football or to provide a magic formula for winning matches. Instead, it demonstrated that the underlying mathematics and the computer code used to solve these problems were sound and reliable. The team verified their results by checking the calculations in multiple ways, ensuring that the computer's gradient checks were accurate and that the solution did not change if the grid used for calculation was made finer.

The ultimate goal of this research is to build a foundation for future tools that could help coaches. By proving that a complex, multi-player tactical problem can be solved with a verified computer algorithm, the researchers have created a reproducible benchmark. This means other scientists can use the same synthetic data and code to test their own ideas. The next step, according to the authors, is to take this framework and feed it real data from actual matches. Only then can the model be calibrated to the messy reality of the sport, where players get tired, the ball moves unpredictably, and opponents react dynamically. For now, the study stands as a rigorous proof that the mathematics of football tactics can be mapped, calculated, and optimized, offering a new lens through which to view the strategic depth of the game.

Technical Summary: Constrained Multi-Agent Optimal Control for Football Tactical Decision Support

1. Problem Statement

This research addresses the prescriptive question of how a football team's collective configuration should evolve over time to achieve specific tactical objectives, given explicit motion constraints. Unlike descriptive or predictive analytics that analyze past events or forecast outcomes, this work formulates tactical planning as a constrained multi-agent optimal control problem.

The core challenge is to determine the optimal velocity controls for $N_p$ players (modeled as states) over a time horizon $T$ on a pitch $\Omega$ . The system must balance competing tactical goals:

Territorial Control: Maximizing influence over specific spatial regions.
Defensive Exposure: Minimizing gaps between linked defenders.
Compactness: Maintaining a cohesive team shape.
Pressing Coverage: Reducing distance to opponent players in high-pressure zones.
Movement Effort: Minimizing the energy cost of velocity changes.
Terminal Accuracy: Reaching a target formation at the end of the horizon.

The problem is subject to physical constraints, including maximum player speed ( $V_{max}$ ) and the requirement that all players remain within the pitch boundaries.

2. Methodology

Mathematical Formulation

The authors model the pitch as a closed rectangle $\Omega = [0, L] \times [0, W]$ . Player positions $x_i(t)$ evolve according to first-order kinematics $\dot{x}_i = u_i$ , where $u_i$ is the velocity control. The opponent is treated as a prescribed path $Y(t)$ rather than an optimizing agent, isolating the single-team control problem.

The objective is to minimize a dimensionless Bolza functional $J(U)$ , which combines:

Running Cost: A weighted sum of normalized tactical losses (territorial, defensive, compactness, pressing, and effort) integrated over time.
Terminal Cost: A penalty for deviation from a target formation $X_{tar}$ .

The spatial control share $P_A(z)$ is modeled using exponential influence fields based on arrival times, normalized against the opponent's influence. All loss functions are smoothed (e.g., using regularization parameters $\kappa, \delta, \rho$ ) to ensure differentiability.

Theoretical Analysis

Existence: The authors prove the existence of an optimal control using the direct method in the calculus of variations, relying on the compactness of the state space and the weak lower semicontinuity of the effort term.
Optimality Conditions: Using Pontryagin's Minimum Principle, the authors derive first-order necessary conditions. This yields a system of state equations, adjoint (costate) equations, and a transversality condition.
Control Law: An explicit interior control law is derived, showing that the optimal velocity is a projection of a gradient term onto the admissible velocity ball $B_{V_{max}}$ .

Numerical Solver

To solve the problem numerically, the authors employ a time-discrete projected-gradient algorithm:

Discretization: The continuous problem is discretized using a piecewise-constant control approximation and the rectangle rule for integration.
Adjoint Method: An exact discrete adjoint is computed backward in time to determine the reduced gradient of the objective function with respect to controls.
Optimization: The algorithm uses Armijo backtracking line search combined with projection onto the feasible set (speed constraints and pitch boundaries).
Verification: The solver is rigorously verified using finite-difference gradient checks, constraint audits, spatial grid sensitivity analysis, and multi-start experiments.

3. Key Contributions

The paper makes six specific methodological contributions:

Unified Formulation: A constrained multi-agent framework that simultaneously encodes territorial, defensive, compactness, pressing, and effort losses as differentiable components.
Theoretical Rigor: Establishment of existence results, smoothness properties, and explicit Pontryagin necessary conditions for the tactical control problem.
Discrete Adjoint & Solver: Derivation of a consistent time-discrete adjoint and a projected-gradient method with Armijo backtracking, including a proof of first-order stationarity for the discrete iteration.
Implementation Fidelity: The provision of an actual implementation of the optimization algorithm, distinguishing the output from simple prescribed interpolations.
Solver Verification: Comprehensive verification via finite-difference gradient checks, constraint auditing, and sensitivity analysis to ensure numerical consistency.
Reproducibility: The release of fully documented synthetic coordinates, parameters, code, and numerical outputs, allowing independent reproduction without proprietary data.

4. Results

Benchmark Performance

Using a synthetic 11-player benchmark on a $105 \times 68$ m pitch over a 12-second horizon:

Objective Improvement: The optimized strategy reduced the objective function value from 0.20935 (feasible linear transition) to 0.18511, an 11.58% improvement. Compared to a static compact baseline (0.29982), the improvement was 38.26%.
Trade-offs: The optimized solution accepted higher movement effort (0.0727 vs. 0.0495 for linear) to achieve significant reductions in pressing loss (0.0269 vs. 0.1140) and territorial loss.
Constraint Satisfaction: All speed constraints were satisfied (max speed exactly 7.0 m/s), and no pitch-boundary constraints became active during the optimization.
Convergence: The solver converged in 88 iterations with a residual of $1.57 \times 10^{-4}$ .

Verification and Sensitivity

Gradient Accuracy: Finite-difference checks showed a maximum relative difference of $3.05 \times 10^{-5}$ between the adjoint gradient and numerical approximations.
Grid Sensitivity: Increasing the spatial quadrature grid from $40 \times 26$ to $120 \times 78$ changed the territorial loss by only 0.135%, confirming the result is not an artifact of coarse discretization.
Multi-Start Robustness: Different initializations (static, linear, random) converged to solutions with a relative spread of only 0.043% in the final objective.

Priority Experiments

The authors tested the model's responsiveness to tactical weights by varying the priority vector:

Pressing-First: Reduced pressing loss further (0.0234) with minimal change in effort.
Defense-First: Reduced defensive loss and movement effort (RMS speed dropped from 1.888 to 1.630 m/s) but increased pressing loss (0.0548).
These results demonstrate that the solver produces coherent, interpretable trade-offs based on the specified weights rather than hard-coded behaviors.

5. Significance and Claims

The paper positions itself as a methodological and computational framework rather than a validated match-prediction system.

Bridging the Gap: The work closes a gap between formulating optimal control problems for football and actually solving them to generate trajectories. It connects abstract tactical indicators to a verified numerical solver.
Reproducibility: By providing synthetic data and code, the authors establish a reproducible basis for future calibration with real tracking data.
Limitations and Scope: The authors explicitly state that the current study uses synthetic data and does not claim predictive validity, causal coaching benefits, or performance in professional matches. The motion model is first-order (ignoring acceleration and biomechanics), and the opponent is prescribed rather than adaptive.
Future Work: The paper suggests that the next logical step is the calibration of tactical weights and initial states using real tracking data, followed by extensions to second-order motion models and differential games for adaptive opponents.

In summary, this paper provides a mathematically rigorous, computationally verified, and fully reproducible framework for generating optimal tactical trajectories in football, serving as a foundation for future data-driven tactical analysis.