Speaker
Description
We study interactive preference elicitation in Multiple Criteria Decision Aiding when the goal is to learn a decision maker’s additive value model from a limited number of sequential pairwise-comparison queries. At each round, the analyst updates the belief about the unknown preference model based on the observed comparison and must select the next pair of alternatives to maximize the informational value of the remaining interactions. To address this problem, we develop a Bayesian preference-learning framework that infers the posterior distribution over the parameters of the additive value model from pairwise comparisons. To make repeated posterior updates computationally tractable during interaction, we approximate Bayesian inference by variational Bayes. We then formulate query selection as a finite-horizon Markov Decision Process, where the state represents the accumulated preference information and determines the current posterior belief over the preference model, transitions are governed by Bayesian posterior updates after each response, and rewards are defined as reductions in posterior uncertainty. The corresponding long-horizon questioning policy is approximated by Monte Carlo Tree Search, which evaluates candidate queries beyond immediate one-step gains. We consider two reward specifications based on posterior uncertainty, one defined in the parameter space and the other in the predictive comparison space. Computational studies on real-world and synthetic MCDA instances show that the proposed Bayesian inference procedure accurately and efficiently recovers preference information, while the MCTS-based questioning policy yields larger reductions in posterior uncertainty than myopic baselines under fixed interaction budgets. The results indicate that combining tractable Bayesian updating with long-horizon planning can improve the effectiveness of sequential pairwise-comparison elicitation in MCDA.
| Keywords | Multiple criteria analysis; Preference elicitation; Preference learning; Variational Bayesian; Monte Carlo Tree Search. |
|---|---|
| Participation in the SNSF Early-Stage Researcher Support and Award Scheme | I am submitting a regular presentation and would like to participate in the SNSF Early-Stage Researcher Support and Award Scheme. |