Reinforcement learning hyperheuristics for optimisationTools Alanazi, Fawaz (2017) Reinforcement learning hyperheuristics for optimisation. PhD thesis, University of Nottingham.
AbstractHyperheuristics are search algorithms which operate on a set of heuristics with the goal of solving a wide range of optimisation problems. It has been observed that different heuristics perform differently between different optimisation problems. A hyperheuristic combines a set of predefined heuristics, and applies a machine learning technique to predict which heuristic is the most suitable to apply at a given point in time while solving a given problem. A variety of machine learning techniques have been proposed in the literature. Most of the existing machine learning techniques are reinforcement learning mechanisms interacting with the search environment with the goal of adapting the selection of heuristics during the search process. The literature on the theoretical foundation of reinforcement learning hyperheuristics is almost nonexisting. This work provides theoretical analyses of reinforcement learning hyperheuristics. The goal is to shed light on the learning capabilities and limitations of reinforcement learning hyperheuristics. This improves our understanding of these hyperheuristics, and aid the design of better reinforcement learning hyperheuristics. It is revealed that the commonly used additive reinforcement learning mechanism, under a mild assumption, chooses asymptotically heuristics uniformly at random. This thesis also proposes the problem of identifying the most suitable heuristic with a given error probability. We show a general lower bound on the time that "every" reinforcement learning hyperheuristic needs to identify the most suitable heuristic with a given error probability. The results reveal a general limitation to learning achieved by this computational approach. Following our theoretical analysis, different reusable and easytoimplement reinforcement learning hyperheuristics are proposed in this thesis. The proposed hyperheuristics are evaluated on wellknown combinatorial optimisation problems. One of the proposed reinforcement learning hyperheuristics outperformed a stateoftheart algorithm on several benchmark problems of the wellknown CHeSC 2011.
Actions (Archive Staff Only)
