Training Greedy Policy for Proposal Batch Selection in Expensive Multi-Objective Combinatorial Optimization

These problems focus on identifying designs, represented as discrete objects like strings or graphs, that optimize multiple Active learning is increasingly adopted for expensive attributes, often requiring substantial resources for accurate multi-objective combinatorial optimization assessment (Ehrgott, 2005; Gómez-Bombarelli et al., 2016; problems, but it involves a challenging subset Stanton et al., 2022; Winter et al., 2019; Mirhoseini et al., selection problem, optimizing the batch acquisition 2021). Active learning frameworks, which iteratively propose score that quantifies the goodness of a batch of candidates and learn from the attributes a batch for evaluation. Due to the excessively evaluated on those candidates, are increasingly employed in large search space of the subset selection problem, these fields due to their query efficiency, which is a critical prior methods optimize the batch acquisition component to handling expensive evaluation costs (Aggarwal on the latent space, which has discrepancies with et al., 2014; Jain et al., 2022; Gruver et al., 2023; Zhu the actual space, or optimize individual acquisition et al., 2023; Agnesina et al., 2023). In active learning, each scores without considering the dependencies round entails an internal problem of selecting a proposal among candidates in a batch instead of directly batch of candidates for querying, formulated by cardinalityconstrained optimizing the batch acquisition.