Differential Evolution for Dynamic Constrained Continuous Optimisation

Abstract

In this thesis, we choose the evolutionary dynamic optimisation methodology to tackle dynamic constrained problems. Dynamic constrained problems represent a common class of optimisation that occur in many real-world scenarios. Evolutionary algorithms are often considered very general search heuristics. Their main advantages (in comparison to problem-specific search methods) are their robustness, flexibility and extensibility, as well as the fact that almost no domain knowledge is required for their implementation and application. Our research is focused on the following areas. In the first part of the thesis, we modify common constraint handling techniques from static domains to suit dynamic environments. We investigate the deficiencies of such techniques and the potential of each method based on the change characteristics of the environment. In the second part, we propose a framework to create benchmarks, since we have observed a lack of benchmarks to evaluate algorithms in dynamic continuous optimisation. Third, we carry out an exhaustive empirical study of diversity mechanisms applied to solve dynamic constrained optimisation problems. Finally, we investigate the integration of a neural network into the evolution process and analyse it’s effectiveness compared to that of popular diversity mechanisms. We address the possibility of integrating such mechanisms with a neural network approach in order to improve the results.