Assessing decision heuristics using machine learning

摘要

In decision making, automatization associated with ‘skill’ contributes to the discrepancies often found between actual and reported use of information. Designing decision support systems (DSS) to assist skilled decision making is complicated by this lack of accessible models. This paper proposes a model of skilled decision making in which tacit heuristics define simplified situations within which skilled decisions are made. Heuristics employed in decision making are identified from qualitative representations of past decisions using machine learning. These heuristics can then be matched with performance criteria to identify conditions under which the user may need supplemental aiding. This use of machine learning to identify ‘buggy’ models of human performance has proved very difficult even for simple tasks such as subtraction. We argue that skilled interaction with real-time systems provides a special case in which the learning problem can he simplified sufficiently to allow user modelling of adequate quality for decision support. Our approach is illustrated in assessments of decision heuristics used by pilots in avoiding intruding aircraft and for operators of a simulated process plant. The ability to adapt decision support functions to complement the decision making processes of skilled users is argued to he essential to providing them effective support Attribute-based machine learning methods are proposed as a feasible mechanism for achieving this goal.