Hyper-Heuristics based on Reinforcement Learning, Balanced Heuristic Selection and Group Decision Acceptance.

In this paper, we introduce a multi-objective selection hyper-heuristic approach combining Reinforcement Learning, (meta)heuristic selection, and group decision-making as acceptance methods, referred to as Hyper-Heuristic based on Reinforcement LearnIng, Balanced Heuristic Selection and Group Decision AccEptance (HRISE), controlling a set of Multi-Objective Evolutionary Algorithms (MOEAs) as Low-Level (meta)Heuristics (LLHs). Along with the use of multiple MOEAs, we believe that having a robust LLH selection method as well as several move acceptance methods at our disposal would lead to an improved general-purpose method producing most adequate solutions to the problem instances across multiple domains. We present two learning hyper-heuristics based on the HRISE framework for multi-objective optimisation, each embedding a group decision-making acceptance method under a different rule: majority rule (HRISE_M) and responsibility rule (HRISE_R). A third hyper-heuristic is also defined where both a random LLH selection and a random move acceptance strategy are used. We also propose two variants of the late acceptance method and a new quality indicator supporting the initialisation of selection hyper-heuristics using low computational budget. An extensive set of experiments were performed using 39 multi-objective problem instances from various domains where 24 are from four different benchmark function classes, and the remaining 15 instances are from four different real-world problems. The cross-domain search performance of the proposed learning hyper-heuristics indeed turned out to be the best, particularly HRISE_R, when compared to three other selection hyper-heuristics, including a recently proposed one, and all low-level MOEAs each run in isolation. • Proposed three selection hyper-heuristics for multi-objective optimisation. • Hyper-heuristics embed reinforcement learning and group decision-making acceptance. • A new quality indicator and two new Late Acceptance variants are presented. • Dynamic control mechanism for running multi-objective evolutionary algorithms. • Best performance across 39 synthetic and real-world multi-objective problem instances. [ABSTRACT FROM AUTHOR]