Introduction
In human-automation teams, such as an AI assisting a radiologist with medical imaging, the standard operating procedure involves the automation handling routine cases while referring ambiguous or difficult tasks to a human expert. However, human cognitive performance is not a static variable; it degrades as mental fatigue accumulates from sustained high task loads.
Existing decision referral algorithms often treat human performance as fixed or only consider immediate workload, ignoring the long-term dynamics of cognitive states. If an AI myopically assigns too many difficult tasks to a human, the operator's fatigue builds up, situational awareness drops, and error rates climb. This paper introduces a dynamic decision referral system that treats human fatigue as a controllable state, optimizing joint decision-making over a long time horizon by incorporating periods of recovery.
Methodology
The authors formulate the automation's decision-making process as a Markov Decision Process (MDP). The state includes the automation's observations of binary classification tasks and the human's current fatigue level. Fatigue is modeled as a controlled Markov process: high task loads increase the probability of transitioning to higher fatigue states, including an "absorbing" state where recovery is impossible, while low task loads allow for recovery.
To overcome the computational complexity of solving this MDP, the authors employ Approximate Dynamic Programming (ADP) with three key innovations:
- Referral Indexing: Proving that for any given task load, the optimal strategy is to refer tasks with the highest "referral indices," balancing automation uncertainty with expected human accuracy.
- Post-Decision States: Simplifying the value function to focus purely on the human's fatigue state.
- Monte Carlo Approximation: Using randomized sampling to approximate expected future costs, ensuring the model remains computationally feasible while converging to the true value function.
Key Insights
- Anticipation beats reaction: Unlike myopic policies that only minimize immediate costs, the ADP policy anticipates fatigue buildup. It proactively reduces task loads before the human reaches a critical fatigue threshold, allowing for cognitive recovery.
- Protecting the human protects the system: The ADP policy may accept slightly higher immediate costs on the automation side (forcing the AI to guess on ambiguous tasks) to prevent catastrophic human errors later in the sequence.
- Decoupling observations: By recognizing that task batches are independent over time, the authors decoupled the massive observation space from the fatigue state space, making the dynamic programming problem solvable.
Key Results
Simulations using Monte Carlo paths demonstrate that the ADP algorithm significantly outperforms baseline myopic strategies over a 10-stage time horizon.
As shown in the fatigue and taskload dynamics, the ADP policy immediately drops the human's task load when fatigue is high, allowing for recovery. In contrast, baseline algorithms maintain high workloads, trapping the human in a high-error state. This proactive management results in a much lower mean cumulative cost over time.
Furthermore, the ADP policy proved robust to model inaccuracies, maintaining lower total costs even when transition probabilities in the fatigue model were perturbed by up to 5% noise.
Limitations and Future Work
The current framework relies on a pre-defined mathematical model of human fatigue dynamics, which may vary significantly between individuals. Future work will involve validating these simulation results with human subject testing to account for the subjective nature of fatigue. Additionally, the authors plan to explore Reinforcement Learning (RL) to allow the automation to learn an individual's fatigue dynamics in real-time and extend the system to multi-class classification problems.