Marathon Race Planning: A Case-Based Reasoning Approach

Barry Smyth, Padraig CunninghamIJCAI 2018 · DOI

TL;DR

This paper introduces a novel Case-Based Reasoning (CBR) system designed to help marathon runners achieve their personal bests. By analyzing past race data, the system predicts an achievable target finish time and generates a personalized pacing strategy, offering a data-driven approach to race planning.

What Problem Does This Paper Address?

Marathon running is a complex endeavor where achieving a personal best (PB) is influenced by numerous factors, including training, race strategy, and even day-to-day conditions. While many runners aim to improve their times, they often struggle with two key aspects: setting a realistic yet challenging target finish time and devising an effective pacing plan to achieve it. A common pitfall for recreational runners is the "positive split," where they run the second half of the race slower than the first, often due to poor pacing or hitting "the wall." This paper aims to address this by providing a system that leverages a runner's own historical performance and the collective experience of similar runners to offer tailored guidance.

Methods

The core of this work is a Case-Based Reasoning (CBR) approach. CBR systems solve new problems by retrieving and adapting solutions from similar past problems (cases). Here, a "case" represents a single runner's experience, specifically a pair of race records: one for a non-personal best (nP B) race and one for their personal best (PB) race.

  1. Case Generation: For each runner with multiple race records, their best-performing race is identified as the PB. This PB record is then paired with each of their other (nP B) race records to create multiple cases. The nPB record serves as the "description" of the case, and the PB record serves as the "solution."
  2. Case Retrieval: When a query runner (with a recent race record) is presented, the system first filters the case base to include only runners with similar finish times and the same gender. Then, it uses a distance-weighted k-Nearest Neighbors (kNN) approach to find the top k most similar cases based on the pacing profiles of the nPB records.
  3. Prediction and Recommendation:
    • Finish-Time Prediction: Three methods are used to predict the runner's achievable PB finish time:
  4. Best PB: Uses the PB time from the single most similar case.
  5. Mean PB: Calculates a weighted average of the PB times from the k most similar cases.
  6. Even PB: Selects the case with the most consistent pacing profile (lowest coefficient of variation) and uses its PB time.
  7. Pacing Plan Recommendation: Three corresponding methods generate pacing plans:
  8. Best Profile: Adapts the pacing profile of the most similar case to the predicted PB time.
  9. Mean Profile: Generates a plan based on the average relative paces from the PB profiles of the k retrieved cases.
  10. Even Profile: Uses the pacing profile of the case with the most even pacing.

Datasets

The evaluation was performed using a dataset of 12,968 race records from the London Marathon spanning the years 2011-2016. This dataset comprised runners who had completed at least three races. The authors used 10-fold cross-validation for evaluation.

Key Insights and Ideas

  • Leveraging Past Success: The fundamental idea is that a runner's past performance, especially when compared to similar runners who achieved PBs, holds valuable predictive power for future performance.
  • Gender Differences in Pacing: The study found that women tend to run more evenly paced races than men, leading to better prediction accuracy and profile similarity for female runners across all strategies.
  • The "Mean" Strategy's Robustness: The Mean PB prediction and Mean Profile recommendation strategies generally performed best, showing lower prediction errors and higher profile similarities, especially as the number of retrieved cases (k) increased. This suggests that averaging insights from multiple similar experiences is more robust than relying on a single best or most even case.
  • "Best" Strategy's Ambition: The "Best" strategy, while aiming for the fastest PB, tended to become overly ambitious as more cases were considered, leading to less accurate predictions for test runners.
  • Ability Matters: Runner ability (indicated by finish times) significantly impacts prediction accuracy. Faster runners benefit from more accurate predictions, while prediction errors increase for slower runners.

Key Results

  • Prediction Accuracy: The Mean strategy achieved the lowest prediction errors, around 4.5% on average for all runners and as low as 4% for women. The Even strategy had a consistent error of ~6%, while the Best strategy's accuracy deteriorated with an increasing number of retrieved cases.
  • Profile Similarity (Recommendation Quality): The Mean strategy also yielded the highest profile similarity, indicating better pacing plan recommendations. Women generally showed higher profile similarity than men.
  • Influence of k: For the Mean strategy, prediction error and profile similarity improved with an increasing number of retrieved cases (k), up to k ≥ 10.
  • Influence of Finish Time: Prediction errors increased with slower finish times for all strategies, but the Mean strategy remained the most accurate. Profile similarity also decreased for slower runners.

Limitations and Future Work

  • Limitations:
    • The "Best" strategy's tendency to predict overly ambitious times as k increases is a notable limitation.
  • The reliance on a single similarity metric (Euclidean distance for pacing profiles) might be a simplification.
  • The dataset is specific to the London Marathon, which might have unique course characteristics.
  • Future Work: The authors suggest further research into more comprehensive treatments of CBR for marathon planning, potentially exploring more sophisticated similarity metrics, incorporating additional physiological data, and testing the approach on different marathon courses.

Summary updated

BibTeX
@inproceedings{smyth2018marathon,
  author = {Barry Smyth and Padraig Cunningham},
  title = {{Marathon Race Planning: A Case-Based Reasoning Approach}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/754},
}
Figure