TL;DR
This paper introduces a novel modeling framework that separates a tennis player's intention (where they aim) from their execution (where the ball actually lands). By using Markov processes and millions of simulated shots, the researchers demonstrate that as execution error increases, the optimal strategy shifts from aggressive "line-painting" to conservative "center-court" play, and that the serve return is the single most impactful shot to optimize.
What Problem Does This Paper Address?
In sports analytics, we often use "value functions" to determine the best move a player can make. However, most models implicitly assume a "chess-like" environment: if you decide to move a piece to a specific square, it goes there with 100% certainty.
Tennis is not chess. A player might intend to hit a screaming winner down the line, but due to physical limitations, the ball lands three feet wide. Traditional models often bake this "execution error" into transition probabilities, making it impossible to tell if a player lost a point because they made a poor tactical choice (bad intention) or simply missed a difficult shot (bad execution). This paper aims to untangle these two factors to understand how a player’s physical precision should dictate their tactical strategy.
Methods
The authors developed a framework using two types of Markov models:
- Markov Reward Processes (MRP): Used for policy evaluation—calculating the win probability of a player's current, empirical strategy across different levels of execution error.
- Markov Decision Processes (MDP): Used for policy optimization—finding the mathematically "perfect" strategy for a given error level.
Key Technical Innovations: * The $\epsilon$ Parameter: The researchers modeled execution error by fitting bivariate Gaussian distributions to ball landing locations. They introduced a controllable scaling parameter, $\epsilon$, to adjust the "spread" of these distributions. $\epsilon = 1$ represents near-perfect execution, while $\epsilon = 13$ was found to represent the average professional player. * State Space Discretization: The court was divided into 84 cells. A "state" was defined by the positions of both players and the shot type (serve, return, or rally). * "Stitched" Simulations: Since shots in their data source were independent, they used a generative framework to "stitch" shots together—using the terminal state of one shot as the starting condition for the next—to simulate entire points.
Datasets
The models were powered by the VON CRAMM framework, which uses an infinite Bayesian Gaussian mixture model trained on:
- 125,000 men’s shots and 80,000 women’s shots from past Australian Opens.
- The framework provided accurate 3D ball trajectories and 2D player tracking data, allowing for the simulation of hundreds of millions of shots to build robust transition probabilities.
Key Insights and Ideas
- The "Conservative Shift": The most striking finding is that "optimal" play is relative to skill. For a player with perfect execution ($\epsilon=1$), the MDP suggests aiming for corners and sidelines. As $\epsilon$ grows, the optimal "aiming points" migrate toward the center of the court to provide a safety buffer against out-of-bounds errors.
- The Backhand Tax: The data shows that execution error on the "ad-side" (typically the backhand for the 90% of players who are right-handed) is significantly more detrimental to win probability than forehand errors.
- The ROI of the Serve Return: If a player can only "optimize" one type of shot in their game, it should be the serve return. The model shows that choosing the optimal intention on a serve return provides a higher boost to win probability than optimizing any other shot type, regardless of the player's execution error.
- Strategy vs. Skill: The authors found that for an average player, being able to choose the "perfect" target for just one or two shots per point is roughly equivalent to having perfect physical execution across their entire current strategy.
Key Results
- Validation: The model predicted a 65.2% win probability for the server at average error levels ($\epsilon=13$), which aligns closely with real-world Grand Slam data (64.2%).
- Aggression Threshold: Aggressive playstyles (aiming for lines) only outperform conservative styles when execution error is very low ($\epsilon < 4$). For the "average" pro ($\epsilon=13$), a conservative strategy actually yields a higher win probability.
- Diminishing Returns: While optimizing every shot (the full MDP) leads to very high win probabilities, the majority of the value is captured within the first 5-6 shots of a point.
Limitations and Future Work
- The "Average" Serve: The current data does not distinguish between first and second serves, instead using an "average" of both. This ignores the tactical nuances of risking a big first serve versus a safe second serve.
- Incoming Shot Difficulty: The state space currently tracks player positions but not the difficulty of the ball being returned (e.g., a high-velocity deep ball vs. a weak short ball).
- Symmetric Opponents: The model assumes the opponent (Player B) always plays with "average" professional execution. Future work could involve a "game theory" approach where both players optimize their strategies simultaneously.
- Expansion to Other Sports: The authors suggest this "controllable error" framework could be applied to any sport with clear intended targets, such as curling, darts, or golf.
Summary updated