Presentation
Inverse Counterfactual for AI-Assisted Decision Support: Enhancing Knowledge Elicitation for Capturing Aircraft Pilot Decisions
DescriptionIntegrating AI into decision-support systems (DSS) for safety-critical tasks like piloting poses challenges for human-machine interaction. DSS must align with the mental model of pilots and provide relevant information. One promising approach, the Cognitive Shadow, addresses this challenge by continuously modelling each operator strategy and issuing alerts whenever its recommendation diverges from the operator decision. We evaluated a new operator-driven knowledge-elicitation procedure: the inverse counterfactual. This procedure aims to enhance human-AI model similarity. After choosing the optimal option, the user modifies one factor, so the second-best option becomes preferable, providing the AI with a contrasting case across the decision boundary.
In a simulated adverse-weather avoidance task, forty-four participants solved 130 scenarios and generated inverse counterfactuals for 20 additional cases. On the final test phase, models trained with these paired cases demonstrated comparable predictive accuracy to conventional models. However, when edits in counterfactual cases were minimal—i.e., closer to the decision boundary— predictive accuracy improved and DSS advice was accepted more frequently. Larger edits degraded performance. The results suggest that well-guided counterfactual interactions may sharpen knowledge elicitation.
The presentation will introduce the methodology, discuss the results, and outline future research aimed at refining the method to better align with human cognitive processes.
In a simulated adverse-weather avoidance task, forty-four participants solved 130 scenarios and generated inverse counterfactuals for 20 additional cases. On the final test phase, models trained with these paired cases demonstrated comparable predictive accuracy to conventional models. However, when edits in counterfactual cases were minimal—i.e., closer to the decision boundary— predictive accuracy improved and DSS advice was accepted more frequently. Larger edits degraded performance. The results suggest that well-guided counterfactual interactions may sharpen knowledge elicitation.
The presentation will introduce the methodology, discuss the results, and outline future research aimed at refining the method to better align with human cognitive processes.
Event Type
Lecture
TimeThursday, October 16th2:10pm - 2:30pm CDT
LocationGrand Hall J
Cognitive Engineering & Decision Making
Similar Presentations