Sustainable Multi-Agent Crowdsourcing via Physics-Informed Bandits
New AI system achieves top reward with only 7.6% workforce utilization, solving a key four-way tension in gig platforms.
A new research paper titled 'Sustainable Multi-Agent Crowdsourcing via Physics-Informed Bandits' introduces a novel AI framework, FORGE, designed to solve a critical four-way tension in gig economy platforms. Authored by Chayan Banerjee, the work addresses the competing demands of allocation quality, workforce sustainability, operational feasibility, and strategic contractor behavior—a dilemma formalized as the Cold-Start, Burnout, Utilisation, and Strategic Agency Dilemma. Existing methods like greedy heuristics or multi-armed bandits fail by causing worker burnout or being operationally infeasible. FORGE reframes the problem by modeling each contractor as a rational agent with its own load-acceptance threshold based on fatigue, turning a standard Restless Multi-Armed Bandit (RMAB) problem into a Stackelberg game.
The core technical innovation is a Neural-Linear UCB allocator that fuses a Two-Tower embedding network with a Physics-Informed Covariance Prior. This prior is derived from offline simulator interactions and simultaneously warm-starts the skill-cluster geometry and the Upper Confidence Bound (UCB) exploration landscape. This provides a geometry-aware belief state from the very first episode, drastically reducing cold-start problems. In simulations over T=200 episodes, the proposed method achieved the highest reward (LRew = 0.555 ± 0.041) among all non-oracle methods while utilizing only 7.6% of the workforce—a combination unattainable by conventional baselines. The system also demonstrated robustness, handling workforce turnover up to 50% and observation noise with a standard deviation (σ) of 0.20, showcasing its potential for real-world deployment on platforms like Amazon Mechanical Turk or Upwork.
- FORGE system solves a four-way tension (quality, sustainability, feasibility, strategy) in crowdsourcing that stumps current methods.
- Neural-Linear UCB allocator achieved top reward (0.555) with just 7.6% workforce use, a feat no baseline could match.
- Demonstrated robustness to 50% worker turnover and significant noise (σ=0.20), making it viable for real platforms.
Why It Matters
This could make gig platforms like Uber or TaskRabbit more sustainable by preventing worker burnout while maintaining service quality and efficiency.