🔬 physics

Evaluating Soccer Player Movements Using the Attacker-Defender Model

This study enhances and validates the attacker-defender motion model using a large dataset of 306 J1 League matches to optimize player-specific parameters and reveal distinct playing styles for attackers and defenders.

Original authors: Takuma Narizuka, Issei Yamazaki

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Takuma Narizuka, Issei Yamazaki

Original paper licensed under CC BY 4.0 (http://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

Imagine a soccer match not just as a game of skill, but as a complex dance between two partners: the person with the ball (the Attacker) and the person trying to stop them (the Defender).

This paper is like a scientific attempt to write the "choreography rules" for that dance. The authors, Takuma Narizuka and Issei Yamazaki, took an existing set of rules (called the Attacker-Defender or AD model) and upgraded them to handle a much bigger stage with more dancers.

Here is a breakdown of what they did, using simple analogies:

1. The Old Map vs. The New GPS

Previously, researchers tried to predict how players moved using a model that was like a rough sketch. It worked okay for small groups of players, but it had two main problems:

It only looked at a few specific types of moves (like only studying how people walk forward).
It was based on a tiny dataset (like trying to understand all of human behavior by watching only 1,500 people).

The authors wanted to upgrade this to a high-definition GPS. They wanted to see if the rules could explain the movements of thousands of players in real, high-stakes games. They used data from 306 professional matches in Japan's top league (J1 League), covering over 31,000 dribbling moments. That's like watching every single step of a marathon runner instead of just a few steps in a park.

2. The Physics of the Dance

The model treats players like cars with engines and brakes.

Resistance (The Brakes): Players naturally slow down if they don't keep pushing.
Goal Force (The Gas): The attacker wants to drive toward the opponent's goal. The defender wants to drive toward their own goal.
Opponent Force (The Steering): The attacker tries to steer away from the defender. The defender tries to steer toward the attacker.

The math calculates the perfect balance between these forces to predict where a player should go.

3. The Big Upgrade: Solving the Puzzle One Piece at a Time

In the old method, the computer tried to solve the movements of the attacker and defender simultaneously. Imagine trying to solve a Rubik's cube while someone else is shaking the table; it's hard and slow.

The authors introduced a new trick: Fix one player, solve the other.

They took the actual path the defender walked and said, "Okay, Defender, you stay exactly here. Now, Attacker, what rules would make you walk exactly where you actually did?"
Then they flipped it: "Okay, Attacker, you stay exactly here. Defender, what rules make you walk where you actually did?"

The Result: This was like switching from solving a tangled knot to untangling two separate strings. It was faster, cheaper to compute, and it worked better. They successfully recreated 88.5% of the dribbling events (up from 85.4% before), meaning their new "GPS" was much more accurate.

4. Discovering the "Four Personalities" of Players

Once they had the accurate data, they realized players don't just move in one way. By looking at the "forces" in their math, they sorted every player into four distinct personality types based on how they moved relative to the goal and the opponent.

For Attackers (The Ball Carriers):

Type A1 (The Safe Driver): Moves toward the goal but keeps a safe distance from the defender. Think of a center-back building up play, calmly passing the ball.
Type A2 (The Backpedaler): Moves away from the goal but toward the defender. Like a player under pressure from behind, trying to shield the ball.
Type A3 (The Daredevil): Moves toward the goal AND directly at the defender. This is the "one-on-one" duel, the high-risk move near the goal.
Type A4 (The Retreat): Moves away from the goal and away from the defender. Like a player accelerating forward to escape pressure or dribbling backward to draw the defender in.

For Defenders (The Chasers):

Type D1 (The Herder): Moves toward their own goal but toward the attacker. Like a defender backing up while trying to push the attacker toward the sideline.
Type D2 (The Retreat): Moves away from the goal and away from the attacker. A pure retreat to avoid getting beaten.
Type D3 (The Sideline Guard): Moves toward the goal but away from the attacker. Dealing with a dribble coming from the side.
Type D4 (The Aggressive Tackle): Moves away from the goal but toward the attacker. This is the high-risk, high-reward move where a defender jumps in to steal the ball from behind.

5. Who Are the Stars?

The paper didn't just look at the math; it looked at the names. They found that certain players appeared most often in specific "personality" zones.

For example, M. Hosoya (a forward) showed up constantly in the defender charts (D1–D4). This makes sense because forwards are often the ones pressing (acting as the "defender" in the model's eyes) when their team doesn't have the ball.
A. Scholz (a defender) was the most frequent "Safe Driver" (A1) for attackers, showing how defenders often act as the ball carriers when building up play.

The Bottom Line

The authors successfully proved that this "physics-based dance model" works on a massive scale. By simplifying how they calculated the moves, they could analyze thousands of real-life soccer moments. They showed that every player has a unique "movement signature" that can be categorized into four distinct styles, giving us a new way to understand the invisible strategies happening on the field.

What's Next?
The authors admit the "GPS" isn't perfect yet. Sometimes the math pushed players to the very edge of the map, suggesting the rules need a little more fine-tuning (like adding penalties for moving too fast or too slow). But for now, they've given us a much clearer picture of how attackers and defenders really interact.

Technical Summary: Evaluating Soccer Player Movements Using the Attacker-Defender Model

Problem Statement
The study addresses the challenge of modeling short-term player trajectories in soccer analytics. While machine learning approaches offer high predictive accuracy, they often lack physical interpretability. Conversely, physics-based models prioritize interpretability but have historically been limited in scope. Specifically, the Attacker-Defender (AD) model proposed by Brink et al. (2023), which describes interactions between a ball carrier and the nearest defender, had previously been explored only with limited parameter ranges and relatively small datasets (1,573 dribbling events). Consequently, the model's applicability to large-scale data and its ability to capture the full spectrum of playing styles remained underexplored.

Methodology
The authors utilized a dataset comprising tracking and event data from 306 matches in the 2023 J1 League season, provided by DataStadium Inc. The raw position data ( $x, y$ ) recorded at 0.04-second intervals were processed using Savitzky–Golay filters and cubic spline interpolation to reduce noise and compute velocities. From this, 31,028 dribbling instances were extracted based on strict criteria: the nearest defender remained constant, possession lasted over 0.5 seconds, and players traveled more than 5 meters.

The core of the study involves the AD model, which defines the motion of an attacker ( $\vec{v}_a$ ) and a defender ( $\vec{v}_d$ ) using coupled differential equations. These equations incorporate a resistance force proportional to velocity and a driving force determined by goal-oriented and opponent-oriented components. The model is governed by six parameters: maximum driving forces ( $f_a, f_d$ ), time constants ( $\tau_a, \tau_d$ ), and weighting factors for opponent interaction ( $\beta_a, \beta_d$ ).

A key methodological innovation is the proposed independent parameter optimization. Unlike the original approach where all six parameters were optimized simultaneously to minimize total trajectory error, this study solves the equations of motion for one player while fixing the opponent's actual trajectory. This allows for the independent estimation of attacker parameters ( $f_a, \tau_a, \beta_a$ ) and defender parameters ( $f_d, \tau_d, \beta_d$ ). Optimization was performed using the COBYLA algorithm with Latin hypercube sampling for initial parameter generation. The study expanded the analysis to all four quadrants of the parameter space (defined by the signs of $f$ and $\beta$ ), whereas previous work had restricted analysis to specific regions.

Key Contributions and Results

Enhanced Optimization Efficiency and Accuracy: The proposed independent optimization method significantly reduced computational costs by lowering the dimensionality of the parameter space. More importantly, it improved the success rate of parameter estimation. Using the improved approach, 88.5% (27,457) of the 31,028 dribbling events met the accuracy criterion (trajectory error $\varepsilon < 0.1$ for both players), compared to 85.4% (26,492) using the original coupled method.
Large-Scale Validation: The study successfully validated the AD model on a dataset nearly 20 times larger than that used in prior research, demonstrating its robustness across a wide range of J1 League scenarios.
Characterization of Playing Styles: By analyzing the optimized parameters across the full four-quadrant space, the authors identified distinct movement patterns and associated them with specific player roles:
- Attackers:
  - Region A1 ( $f_a > 0, \beta_a > 0$ ): Movement toward the opponent's goal and away from the defender. Dominated by center backs involved in build-up play.
  - Region A2 ( $f_a < 0, \beta_a > 0$ ): Movement away from the goal but toward the defender, often seen when maintaining possession under rear pressure.
  - Region A3 ( $f_a < 0, \beta_a < 0$ ): Movement toward both the goal and the defender, corresponding to direct one-on-one confrontations near scoring chances.
  - Region A4 ( $f_a > 0, \beta_a < 0$ ): Movement away from both the goal and the defender, observed in forward acceleration under pressure or backward dribbles.
- Defenders:
  - Region D1 ( $f_d > 0, \beta_d > 0$ ): Movement toward the defending goal and the attacker, typical of pressing actions by center forwards.
  - Region D2 ( $f_d < 0, \beta_d > 0$ ): Movement away from both the goal and the attacker, representing defensive retreats.
  - Region D3 ( $f_d < 0, \beta_d < 0$ ): Movement toward the goal but away from the attacker, common when dealing with side dribbles.
  - Region D4 ( $f_d > 0, \beta_d < 0$ ): Movement away from the goal but toward the attacker, indicating high-risk challenges from behind.

Significance and Claims
The paper claims that by expanding the range of analyzed parameters and utilizing a larger dataset, the study provides new, quantitative insights into the distinct playing styles of attackers and defenders. The authors assert that the independent optimization method not only improves the efficiency of model fitting but also increases the volume of accurately reproduced trajectories, thereby enabling more robust player characterization.

The study concludes modestly, noting that while the model successfully reproduces trajectories, many optimized parameters approached the boundaries of the defined parameter space. This suggests a need for future refinement of the error function, potentially by incorporating penalty terms based on player velocity and acceleration. Future work is also proposed to analyze specific contexts (e.g., field location) and player roles in greater detail.