Séminaire de Lars Kotthoff
The Shapley Value & the Temporal Shapley Value for Algorithm Performance Analysis
27 nov. 2025 - 14:00It is surprisingly difficult to quantify an algorithm's contribution to the state of the art. Reporting an algorithm's standalone performance wrongly rewards near-clones while penalizing algorithms that have small but distinct areas of strength. Measuring an algorithm's marginal contribution is better, but penalizes sets of strongly correlated algorithms, thereby obscuring situations in which it is essential to have at least one algorithm from such a set. Neither of these measures takes time into account, penalizing algorithms that are no longer state-of-the-art, but were when they were introduced. In this talk, I will argue that contributions should be analyzed via a measure drawn from coalitional game theory, the Shapley value, and its time-sensitive cousin, the temporal Shapley value. The temporal Shapley Value maintains the desirable properties of the Shapley Value, but allows to take the time algorithms were introduced into account, within the context of the state of the art at the time. These measures characterize the contribution of an algorithm fairly and can yield insight into a research community's progress over time.
The proposed measures are applied to the famous quicksort algorithm and algorithms from the AI areas of satisfiability and constraint programming. The results illustrate the benefits of the Shapley Value and its temporal cousin. Our measures give insights into how the field has developed over time and were used in the 2021 AI Index to assess the progress of AI.