Fatigue-Aware Learning to Defer via Constrained Optimisation

2026-04-01

AI Summary

This work addresses a critical limitation in existing learning-to-defer approaches for human-AI collaboration: the assumption of constant human expert performance, which neglects performance degradation due to psychological fatigue. To remedy this, the study introduces, for the first time, a dynamic mental fatigue model into the learning-to-defer framework. The authors formulate the collaborative decision-making process as a constrained Markov decision process (CMDP) that jointly models human and AI states, and optimize the deferral policy using a PPO-Lagrangian algorithm. The proposed method enables zero-shot generalization across experts exhibiting diverse fatigue dynamics and establishes FA-L2D, a new benchmark encompassing multiple fatigue patterns. Experiments demonstrate that the approach significantly outperforms state-of-the-art methods across several datasets, achieving superior accuracy—particularly in mid-coverage regimes—compared to purely AI-driven or human-only decision systems.

Abstract

Learning to defer (L2D) enables human-AI cooperation by deciding when an AI system should act autonomously or defer to a human expert. Existing L2D methods, however, assume static human performance, contradicting well-established findings on fatigue-induced degradation. We propose Fatigue-Aware Learning to Defer via Constrained Optimisation (FALCON), which explicitly models workload-varying human performance using psychologically grounded fatigue curves. FALCON formulates L2D as a Constrained Markov Decision Process (CMDP) whose state includes both task features and cumulative human workload, and optimises accuracy under human-AI cooperation budgets via PPO-Lagrangian training. We further introduce FA-L2D, a benchmark that systematically varies fatigue dynamics from near-static to rapidly degrading regimes. Experiments across multiple datasets show that FALCON consistently outperforms state-of-the-art L2D methods across coverage levels, generalises zero-shot to unseen experts with different fatigue patterns, and demonstrates the advantage of adaptive human-AI collaboration over AI-only or human-only decision-making when coverage lies strictly between 0 and 1.

Problem

• Learning to Defer
• human-AI cooperation
• fatigue
• human performance degradation
• workload

Innovation

• Learning to Defer
• Fatigue Modeling
• Constrained Markov Decision Process
• Human-AI Collaboration
• PPO-Lagrangian