Contest Design with Uncertain Performance and Costly Participation

Priel Levy, David Sarne, Igor RochlinIJCAI 2017 · DOI

TL;DR

This paper analyzes contest design when participants' performance is uncertain and participation incurs a cost. The authors compare two contest formats – parallel (simultaneous decisions) and sequential (decisions made in order, with knowledge of previous outcomes) – to understand how an organizer can best incentivize participation and maximize the quality of the winning contribution. They find that the optimal contest format can surprisingly shift based on subtle changes in participation costs and prize values, even revealing counter-intuitive properties in sequential contests where participants' expected profits are independent of their position.

What Problem Does This Paper Address?

Many real-world scenarios involve contests where participants cannot directly control their performance once they've decided to enter, and their success depends on subjective evaluation, luck, or factors beyond their influence. Think of submitting headshots for an acting role, entering a photography contest, or even applying for academic positions. In these situations, the primary strategic decision for a participant is whether to incur the cost of participation (time, effort, money) for a chance at a prize. Organizers, on the other hand, want to design contests that attract high-quality contributions. This paper tackles the problem of how an organizer should design such contests, specifically by comparing two fundamental structures: parallel contests (everyone decides simultaneously) and sequential contests (participants decide in turn, observing previous outcomes). The core challenge is to understand how these different structures, coupled with uncertain performance and participation costs, affect participant behavior and the organizer's objective of maximizing the best outcome.

Methods

The paper employs a game-theoretic approach to model and analyze the contest dynamics. They define a model with a contest organizer and a set of heterogeneous potential contestants. Each contestant faces a binary choice: participate (incurring a cost $c_i$) or not. Their performance is determined by a probabilistic function $f_i(x)$, and the organizer offers a prize $M$ to the highest performer.

The key methodological contributions are:

  • Comparative Analysis of Parallel vs. Sequential Contests: The paper provides a formal analysis of both parallel (simultaneous) and sequential (ordered, with information) contest designs. This comparison is central to understanding the trade-offs.
  • Bayesian Nash Equilibrium (BNE): For the parallel contest, the analysis focuses on finding BNE strategies, where each agent's strategy is a best response to the others' strategies, considering the uncertainty.
  • Subgame Perfect BNE: For the sequential contest, the analysis uses subgame perfect BNE, which accounts for optimal play at every stage of the game, given the information available at that stage.
  • Homogeneous Agent Analysis: A significant portion of the analysis focuses on the special case where all agents are homogeneous (same participation cost, same performance distribution). This simplifies the analysis and reveals interesting general properties.
  • Characterization of Equilibrium Strategies: The paper derives conditions for pure and mixed strategy equilibria in parallel contests and threshold-based strategies in sequential contests.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Organizer's Preference is Dynamic: The choice between a parallel and sequential contest is not fixed. The organizer's preference can shift dramatically based on subtle changes in parameters like participation costs and prize values.
  • Counter-Intuitive Sequential Equilibrium: In sequential contests with homogeneous agents, participants' strategies and expected profits are surprisingly independent of their position in the sequence and the number of participants. This suggests a level of fairness that might not be immediately obvious.
  • Threshold Strategies in Sequential Contests: Agents in sequential contests adopt a threshold strategy: they participate only if their observed performance (or the best observed so far) exceeds a certain level. This is analogous to optimal stopping problems.
  • Homogeneity Simplifies but Reveals Surprises: While homogeneity simplifies analysis, it leads to counter-intuitive results, such as the independence of profit from position in sequential contests.
  • "Tragedy of the Commons" in Homogeneous Sequential Contests: Even when participation is beneficial for the agents collectively, the equilibrium might lead to no one participating if the prize is not sufficiently high relative to costs, mirroring a "tragedy of the commons" scenario.

Key Results

  • Parallel Contest Equilibria: Proposition 1 characterizes BNE for parallel contests with homogeneous performance distributions. It identifies conditions for all agents participating, all abstaining, or mixed strategies.
  • Sequential Contest Equilibrium: Theorem 1 describes the best-response strategy for an agent in a sequential contest, which is based on a participation threshold.
  • Homogeneous Parallel Contest: Theorem 2 characterizes the BNE for homogeneous parallel contests, showing pure strategy equilibria when $c/M \ge 1/k$ or $c/M \le 1/(k-1)$, and mixed strategies otherwise. The participation probability $p$ is derived.
  • Homogeneous Sequential Contest: Theorems 3 and 4 characterize the BNE for homogeneous sequential contests. Notably, all agents use the same threshold $r$, and all achieve the same expected profit, regardless of their position. This profit is given by $M(F(r))^k / k$ (for the case where not all agents abstain).
  • Method Comparison: Proposition 2 provides conditions for when the parallel contest dominates the sequential one for the organizer, and vice versa, highlighting shifts in preference based on the ratio of participation cost to prize value ($c/M$) and the number of agents ($k$).

Limitations and Future Work

  • Prize Awarding Assumption: The model assumes the prize is always awarded. The authors note that extending this to a scenario where the prize might not be awarded if no one participates is straightforward by adjusting the profit calculation.
  • Organizer's Utility: The current model focuses on the organizer maximizing expected performance. A more complex multi-attribute utility function for the organizer, balancing performance and prize costs, could be explored.
  • Future Directions: The authors suggest several avenues for future research:
  • Multi-prize allocation: Designing contests with multiple prizes to further incentivize participation.
  • Sequencing algorithms: Developing algorithms to optimize sequential contest design for heterogeneous agents and maximize profit.
  • Prize-setting algorithms: Focusing on algorithms that maximize social welfare.
  • Exploring different orderings: The current sequential analysis assumes a fixed order; exploring how different orderings impact outcomes is a natural extension.

Summary updated

BibTeX
@inproceedings{levy2017contest,
  author = {Priel Levy and David Sarne and Igor Rochlin},
  title = {{Contest Design with Uncertain Performance and Costly Participation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/43},
}
Figure