Trade-Offs Between Information and Crowding in Sequential Decisions (Student Abstract)

Daniel Hall, Jon M. KleinbergAAAI 2025 · DOI

Introduction

In social and economic systems, individuals often make decisions by observing the actions of those before them. Classic models of "informational cascades" suggest that people eventually ignore their private information to follow the crowd—a phenomenon seen in everything from choosing a busy restaurant to stock market bubbles. However, these models typically assume that the reward for a correct choice is fixed regardless of how many people choose it.

In reality, many rewards are finite and must be shared. This introduces the force of crowding: as more people choose the "correct" option, the individual payoff diminishes. This paper explores the tension between seeking information (following others to be right) and avoiding crowding (choosing the less popular option to maximize share).

The Model: Information vs. Shared Rewards

The authors formalize this trade-off using a modified sequential urn experiment. An urn is either "Mostly Blue" (MB) or "Mostly Red" (MR) with equal probability. Each of $n$ players sequentially draws a marble, observes all previous draws (public signals), and makes a guess.

The critical departure from standard models is the payoff structure:

  • If a player guesses correctly, they receive a share of a fixed total reward $n$. Specifically, if $k$ players are correct, each receives $n/k$.
  • If a player guesses incorrectly, they receive zero.

To solve this game, the authors developed a polynomial-time algorithm using dynamic programming to identify equilibrium strategies, allowing for the analysis of complex behaviors in large groups.

Strategic Non-Conformity and the First Player's Dilemma

One of the most striking findings is that agents may intentionally choose the option less likely to be correct to avoid future crowding. This "strategic non-conformity" is most evident in the first player's decision.

Figure 1: Decision boundaries for the first player. White regions indicate where the player follows their signal; black regions indicate strategic non-conformity to avoid crowding.

As shown in Figure 1, if the signal strength $p$ is relatively low, the first player might guess against their own signal. By choosing the "wrong" color initially, they may discourage subsequent players from following them, thereby securing a larger portion of the reward if their "unlikely" guess turns out to be correct. For instance, at $n=3$, if $p < 0.585$, the first player will strategically deviate from their signal.

Predictive Crowding and Convergence

The model reveals that agents are forward-looking; they don't just react to current crowding but anticipate how their choice will influence the decisions of those yet to move. This leads to "non-monotonic" behavior: an increase in evidence for an option might actually make it less attractive if it signals that the option will become over-crowded by the end of the game.

Despite these strategic deviations, the authors prove that information eventually dominates.

  • Theorem (Convergence): After $O(\log n)$ signals, the informational advantage becomes so overwhelming that the cost of crowding is secondary. The probability that players converge on the correct answer approaches 1 as the number of players $n$ grows.

Conclusion

This research highlights the complex dynamics that emerge when rewards are shared. While the fear of crowding can lead to counter-intuitive early-stage behavior and strategic "alternation" between choices, the collective accumulation of information eventually leads the group to the truth. Future work remains to characterize the exact patterns of alternation and to explore the impact of private signals on this information-crowding trade-off.

Summary updated

BibTeX
@inproceedings{hall2025trade,
  author = {Daniel Hall and Jon M. Kleinberg},
  title = {{Trade-Offs Between Information and Crowding in Sequential Decisions (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35261},
}

Loading related content...

Figure