BONUS! Maximizing Surprise

Zhihuan Huang, Yuqing Kong, Tracy Xiao Liu, Grant Schoenebeck, Shengwei Xu

TL;DR

This paper provides a mathematical framework for determining the optimal "bonus" points in the final round of a competition to maximize audience excitement. By modeling audience beliefs as a Beta distribution and defining excitement through the lens of information-theoretic "surprise," the authors prove that the ideal bonus often scales with the "expected lead" of the stronger player.

What Problem Does This Paper Address?

Why do games like Quidditch, IndyCar racing, or even WeChat quiz games double or triple the points in the final round? Intuitively, it’s to keep the audience on the edge of their seats. If one player is far ahead, the final round becomes a formality, and viewership drops.

The authors formalize this intuition using the concept of surprise: the absolute change in the audience's belief about who will win from one round to the next. The challenge is that calculating the optimal bonus for an $n$-round game is computationally expensive ($O(n^3)$ via naive backward induction) and depends heavily on what the audience already knows about the players. The paper seeks to find a general, efficient way to calculate the bonus that maximizes the total expected surprise over the entire competition.

Methods

The authors model a two-player, $n$-round competition where the first $n-1$ rounds are worth 1 point each, and the final round is worth a bonus $x$. They assume each player has a constant winning probability $p$, but the audience doesn't know $p$ exactly. Instead, the audience has a prior belief modeled by a Beta distribution $Be(\alpha, \beta)$.

The core of their approach is a Main Technical Lemma. They demonstrate that for Beta priors, the ratio of expected surprise between any two consecutive rounds (up to the penultimate round) is a constant that depends only on the round number and the prior parameters, not on the final bonus $x$.

This insight allows them to reduce a complex multi-round optimization problem into a much simpler trade-off between just two terms: the surprise generated in the final round versus the surprise in the penultimate round. This reduction leads to an $O(n)$ algorithm for the general case and closed-form solutions for specific scenarios.

Datasets

Not discussed in the paper. As a theoretical work in computer science and game theory, the paper relies on mathematical proofs and numerical simulations rather than empirical datasets.

Key Insights and Ideas

  • The "Expected Lead" Intuition: In cases where the audience is fairly certain about the players' relative abilities (the "certain case"), the optimal bonus is approximately equal to the "expected lead"—the number of points the weaker player is expected to be trailing by before the final round. This ensures the final round remains "pivotal."
  • Surprise as a Martingale: The audience's belief curve is a Doob martingale, meaning their current belief is the best predictor of their future belief. Surprise is the "path length" of this belief journey.
  • Skewness vs. Uncertainty:
    • Skewness: If the match-up is uneven (high skewness), you need a larger bonus to give the underdog a meaningful chance to flip the outcome in the end.
  • Uncertainty: In symmetric match-ups (where neither player is a clear favorite), higher uncertainty about their actual skill levels also calls for a larger bonus. This is because uncertainty often leads to lopsided scores early on, which a large bonus can "correct" for the sake of drama.
Figure 2: Belief curves with low/high overall surprise

Key Results

  • Computational Efficiency: The authors developed an $O(n)$ algorithm to find the optimal bonus for any general Beta prior, a significant improvement over the $O(n^3)$ brute-force approach.
  • Asymptotic Behavior: For very large $n$, the optimal bonus ratio $\mu = x/n$ converges to a specific solution of a function $G(\mu) = 0$.
  • Symmetric Case: When the prior is symmetric ($\alpha = \beta$), the optimal bonus has a clean, closed-form solution. If the match is perfectly even and $n$ is odd, the optimal bonus is simply 1 (no extra points needed).
  • Visualizing the Trade-offs: Figure 1 illustrates how the optimal bonus ratio $\mu^*$ changes based on the skewness and uncertainty of the prior. As skewness increases (moving right on the x-axis), the required bonus grows significantly.
Figure 1: Optimal bonus in asymptotic case: for all α,β ≥1, we use α−β

Limitations and Future Work

  • The "End-Effect": The model assumes all surprise is weighted equally. However, psychological research suggests audiences value surprise more at the end of an event (the recency effect). Future work could incorporate time-weighting into the surprise function.
  • Static Bonus Structure: The paper only looks at a bonus in the final round. Real-world games might benefit from escalating point values (e.g., 1, 2, 3...) or bonuses that trigger based on previous results.
  • Variable Game Lengths: The current model assumes a fixed $n$ rounds. Games like Quidditch, where the end is triggered by an event (catching the Snitch), require a more complex stochastic model.
  • Field Validation: The authors suggest that their theoretical predictions should be tested via field experiments, such as AB testing different scoring rules in e-sports or mobile games.

Summary updated

Abstract

Multi-round competitions often double or triple the points awarded in the final round, calling it a bonus, to maximize spectators' excitement. In a two-player competition with $n$ rounds, we aim to derive the optimal bonus size to maximize the audience's overall expected surprise (as defined in [7]). We model the audience's prior belief over the two players' ability levels as a beta distribution. Using a novel analysis that clarifies and simplifies the computation, we find that the optimal bonus depends greatly upon the prior belief and obtain solutions of various forms for both the case of a finite number of rounds and the asymptotic case. In an interesting special case, we show that the optimal bonus approximately and asymptotically equals to the "expected lead", the number of points the weaker player will need to come back in expectation. Moreover, we observe that priors with a higher skewness lead to a higher optimal bonus size, and in the symmetric case, priors with a higher uncertainty also lead to a higher optimal bonus size. This matches our intuition since a highly asymmetric prior leads to a high "expected lead", and a highly uncertain symmetric prior often leads to a lopsided game, which again benefits from a larger bonus.
BibTeX
@inproceedings{huang2021bonus,
  author = {Zhihuan Huang and Yuqing Kong and Tracy Xiao Liu and Grant Schoenebeck and Shengwei Xu},
  title = {{BONUS! Maximizing Surprise}},
  booktitle = {The Web Conference},
  year = {2021},
  doi = {10.1145/3485447.3512049},
}
Figure