Abstract
Most multimedia crowdsourcing models assume workers are "fully rational" agents who always make mathematically optimal decisions. However, real-world human behavior often deviates from this ideal due to limited reasoning depth. This paper introduces a framework based on Cognitive Hierarchy (CH) theory to model "bounded rationality" in crowdsourcing. By analyzing a two-stage game where a requester sets rewards and quality requirements, the authors show that accounting for human cognitive limits allows platforms to optimize strategies and significantly boost profits compared to traditional Nash Equilibrium models.
Introduction
In the domain of multimedia crowdsourcing—such as hiring individuals to capture high-resolution imagery of specific urban landmarks—the success of a platform hinges on the interaction between the requester's incentives and the workers' task selection. Traditional game-theoretic approaches typically rely on the Nash Equilibrium (NE), assuming all participants are perfectly rational. This research argues that such assumptions are unrealistic. Instead, it applies Cognitive Hierarchy theory to better reflect the "bounded rationality" of human workers, who often think only a few steps ahead.
Problem Formulation
The authors model the crowdsourcing environment as a two-stage game involving a central platform, a requester, and a pool of potential workers.
In Stage I, the requester determines the vector of rewards $R = [R_1, R_2, \dots, R_M]$ and quality requirements $Q = [Q_1, Q_2, \dots, Q_M]$ for $M$ different tasks. In Stage II, workers evaluate these parameters and decide which task to perform, if any, based on their individual skill levels and the perceived competition.
Methodology
The study compares two behavioral frameworks:
- Fully Rational (FR) Benchmark: Workers are assumed to reach a Nash Equilibrium. Since finding the NE for heterogeneous workers is computationally expensive (NP-hard), the authors developed a Greedy Randomized Task Set Search Algorithm to approximate optimal requester strategies.
- Bounded Rational (BR) Model: Utilizing Cognitive Hierarchy theory, workers are categorized by their reasoning depth ($k$). Level-0 workers choose tasks randomly; Level-1 workers assume others are Level-0 and respond optimally; Level-$k$ workers assume a distribution of lower-level thinkers.
A key theoretical contribution is the proof of convergence: as the average cognitive level ($\tau$) of the population increases toward infinity, the BR model's equilibrium converges to the traditional Nash Equilibrium.
Key Insights and Results
The research yields several counter-intuitive findings regarding platform profitability:
- The Profit of Imperfection: Requesters can often extract higher profits from boundedly rational workers than from fully rational ones. Because BR workers may "over-compete" for certain tasks, the requester can maintain high participation even with lower rewards.
- The Saturation Gap: While profits in FR models tend to saturate as the worker population grows, the BR model suggests that profit can increase almost unboundedly. This is driven by the persistent presence of low-level thinkers who participate regardless of complex strategic considerations.
- Strategic Reward Reduction: The simulations identified a "threshold" population size. Once the worker pool is sufficiently large, a requester can drastically reduce rewards to the minimum viable cost, relying on the less strategic segments of the population to fulfill tasks.
- Heuristic Efficiency: The proposed Greedy Randomized algorithm achieved results nearly identical to exhaustive searches but with significantly lower computational complexity.
Conclusion
By integrating quality requirements and bounded rationality into incentive design, this work provides a more realistic and profitable roadmap for multimedia crowdsourcing platforms. Future research aims to relax the assumption of complete information regarding worker costs and to validate these theoretical models through large-scale field trials to estimate real-world cognitive levels across different platforms.
Summary updated