Parameterized Heuristics for Incomplete Weighted CSPs with Elicitation Costs

Weighted Constraint Satisfaction Problems (WCSPs) are an elegant paradigm for modeling combinatorial optimization problems. A key assumption in this model is that all constraints are specified or known a priori, which does not hold in some applications where constraints may encode preferences of human users. Incomplete WCSPs (IWCSPs) extend WCSPs by allowing some constraints to be partially specified, and they can be elicited from human users during the execution of IWCSP algorithms. Unfortunately, existing approaches assume that the elicitation of preferences does not incur any additional cost. This assumption is unrealistic as human users are likely bothered by repeated elicitations and will refuse to provide an unbounded number of preferences. Therefore, we propose the IWCSP with Elicitation Cost (IWCSP+EC) model, which extends IWCSPs to include elicitation costs, as well as three parameterized heuristics that allow users to trade off solution quality for fewer elicited preferences and faster computation times. They provide theoretical quality guarantees for problems where elicitations are free. Our model and heuristics thus extend the state of the art in constraint reasoning to better model and solve agent-based applications with user preferences.