A preference elicitation framework for automated planning. (1st December 2022)

• Record Type: Journal Article
• Authors: Mantik, Sheryl; Li, Minyi; Porteous, Julie
• Is Part Of: Expert systems with applications. Volume 208(2022)
• Publication Date: 2022-12-01
• Subjects: Preference learning -- Automated planning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118014
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Many automated planning systems allow users to specify their preferences in order to generate plans that are of high quality according to specified preferences. However, specifying preferences upfront can be time-consuming and difficult for users as their preferences can be complex, not be known in advance, or largely incomplete, especially in complex and large applications. Although many preference representation models and elicitation methods have been proposed in the literature, their adaptation to planning is non-trivial and work in this area has been somewhat limited. In this work, we present a preference elicitation framework for automated planning. This framework allows for interaction with a user through a limited number of low burden comparison type queries and for subsequent learning of a preference relation predictor based on the user's responses. To help the user explore possible planning solutions and to assist them to easily discover and articulate their preferences, the framework prompts the user with planning solutions for small problems and attempts to learn the user's preference through a variety of small planning problems from a particular domain. The learned preferences can then be generalised to infer preferences for planning solutions to larger planning problems. We evaluate the framework using a variety of different benchmark planning domains. The results suggest that our framework can efficiently learn user preferences when provided with a limited number of preference queries in small problems, and generalise the user preferences with high accuracy to larger planning problems. Highlights: We present a preference elicitation framework for automated planning. It allows users to explore small and simple planning problems to elicit preferences. It interacts with users through low burden comparison type queries. It then learns from user query responses using machine learning techniques. Our findings suggest that our framework can efficiently learn from user preferences.