Abstract:
Hyper-heuristic is one of the optimization approaches that are devised with a high degree of abstraction that enable them to work with any area or class of problems and thus are further mostly appropriate than specialized heuristic and meta-heuristic methods. Hyper-heuristics are intended to solve the heuristic selection and generation problem rather than to solve a particular problem in the real world. Hyper-heuristics can therefore be seen as methods for optimising the operations of an optimization process that finds a good solution whenever a problem comes in a new instance of it. This method has proved successful in most of the cases as it improves search performance and decreases the load connected with tailoring meta-heuristics which are often necessary for solving new problems. Selection and generation methodologies are the most common hyper-heuristics in the literature. In this thesis, the hypothesis is tested that selection hyper-heuristic can be applied in a competitive manner to the multi-objective Automatic Design Space Exploration (ADSE) problem of the embedded system. Even though in the literature, many multi-objective meta-heuristic and meta optimization methods have been proposed for the optimization problem in Design Space Exploration (DSE) for embedded systems, however, solution based on selection hyper-heuristics is silent in literature. This work investigates how selection hyper-heuristic affects the quality of the exploration and its runtime and proposes three different selection hyper-heuristics that lead to better results compared to the existing meta-heuristics. This thesis explores and analyses three different selection hyper-heuristics algorithms in which two are no learning selection methods and one is a learning selection method. These have been combined with two deterministic, non¬stochastic move acceptance methods and two non-deterministic, stochastic move acceptance methods. All algorithms are applied to solve the DSE problem of embedded systems where selective hyper-heuristics is shown to be very effective at solving this difficult problem. The newly proposed heuristics are shown to produce improved results as compared to existing meta-heuristics.