Towards Automatic Controller Design using Multi-Objective Evolutionary Algorithms

In order to design the controllers of tomorrow, a need has risen for tools that can aid in the design of these. A desire to use evolutionary computation as a tool to achieve that goal is what gave inspiration for the work contained in this thesis. After having studied the foundations of evolutionary computation, a choice was made to use multi-objective algorithms for the purpose of aiding in automatic controller design. More specifically, the choice was made to use the Non-dominated Sorting Genetic Algorithm II (NSGAII), which is one of the most potent algorithms currently in use, as the foundation for achieving the desired goal. While working with the algorithm, some issues arose which limited the use of the algorithm for unknown problems. These issues included the relative scale of the used fitness functions and the distribution of solutions on the optimal Pareto front. Some work has previously been done in this area using methods based on relative angles, utility functions, and projections and that work is what is extended in this thesis in order to cover a wider range of problems. This allows the NSGA-II to be transformed into a "black-box" optimization tool, which can be used for automatic controller design. However, because the field of evolutionary computation is relatively unknown in the field of control engineering, this thesis also includes a comprehensive introduction to the basic field of evolutionary computation as well as a description of how the field has previously been used for solving a variety of issues in control engineering.