Browsing by Author "Malan, Maria Magdalena"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemParticle swarm optimisation: an algorithm using support vector classification based constraint approximations.(Stellenbosch : Stellenbosch University, 2019-04) Malan, Maria Magdalena; Venter, Gerhard; Stellenbosch University. Faculty of Engineering. Dept. of Mechanical and Mechatronic Engineering (CRSES)ENGLISH ABSTRACT: Particle swarm optimisation (PSO) is by its nature an unconstrained global optimisation method, which must be, and has been, adapted in order to be capable of solving constrained optimisation problems. PSO, along with other similar metaheuristics, struggles when the global optimum is located on the boundary of the feasible region, which is common in most real-world problems, and when there are equality constraints. The aim was to develop a method for improving PSO’s ability to solve these types of constrained optimisation problems. The view that was taken was that the problem with finding global optima on the boundary is the lack of knowledge about the boundary and that deliberately encouraging the exploration of the boundary is critical for having the swarm discover these optima or having the swarm discover them in fewer iterations. To address the lack of knowledge of the boundary, support vector classification (SVC) and the data points that are already evaluated by the swarm were used to create models of the feasibility boundary. The new knowledge that these SVC models provide is used to encourage the particles’ exploration of the boundary, in order to increase the likelihood of locating the global optimum. A thorough literature review is presented to place the concepts into their proper context, present related or similar work and provide the necessary background knowledge. The reasoning behind the concepts that were created and their implementation is laid out. Four concepts with several variations were designed and evaluated through testing on a large set of test problems. The concepts were considered to be additions to a baseline PSO algorithm, and the impact their addition had was evaluated relative to it. The concepts were assessed while mimicking the SVC’s predictions, thus assuming 100% model accuracy, and then with the full classifier attached. Overall, several of the concepts provided significant reductions in the number of iterations that were required on many of the test problems. There were also clear improvements on some of the simpler equality constrained problems. Some problems or challenges are explained, and suggestions are made for improving any future implementations. The most generally promising concept algorithm shifts particles for which the position rule’s new position would entail a move from feasible to infeasible region, back to the approximate point the particle would have to cross the feasibility boundary. The intended application is to optimisation problems where the evaluation of the objective function and constraints significantly dominates the computation time, as in larger engineering design problems with simulations. For these problems the computational overhead that is introduced by the creation and automated tuning of the SVC models could potentially be negligible relative to the overall decrease associated with the reduction in the number of function evaluations required.