Q-learning based hyper-heuristic for scheduling system self-parameterization

Optimization in current decision support systems has a highly interdisciplinary nature related with the need to integrate different techniques and paradigms for solving real-world complex problems. Computing optimal solutions in many of these problems are unmanageable. Heuristic search methods are known to obtain good results in an acceptable time interval. However, parameters need to be adjusted to allow good results. In this sense, learning strategies can enhance the performance of a system, providing it with the ability to learn, for instance, the most suitable optimization technique for solving a particular class of problems, or the most suitable parameterization of a given algorithm on a given scenario. Hyper-heuristics arise in this context as efficient methodologies for selecting or generating (meta) heuristics to solve NP-hard optimization problems. This paper presents the specification of a hyper-heuristic for selecting techniques inspired in nature, for solving the problem of scheduling in manufacturing systems, based on previous experience. The proposed hyper-heuristic module uses a reinforcement learning algorithm, which enables the system with the ability to autonomously select the meta-heuristic to use in optimization process as well as the respective parameters. A computational study was carried out to evaluate the influence of the hyper-heuristics on the performance of a scheduling system. The obtained results allow to conclude about the effectiveness of the proposed approach.