An adaptively switching iteration strategy for population based metaheuristics / Nor Azlina Ab. Aziz

Nor Azlina, Ab. Aziz (2017) An adaptively switching iteration strategy for population based metaheuristics / Nor Azlina Ab. Aziz. PhD thesis, University of Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (1576Kb)
Preview	PDF (Thesis PhD) Download (9Mb) | Preview

Abstract

Population-based metaheuristics are iterative procedures that search for an optimal solution through exploration of the search space and exploitation of information by a group of search agents. The iteration strategy determines how the procedures are executed with respect to the population. Two types of iteration strategies are traditionally available. The first type which is the most commonly adopted strategy is the synchronous update. In the synchronous update, all the search procedures are executed as a group. The entire population needs to complete a particular procedure first before another procedure can be executed. The second type of traditional iteration strategy available is the asynchronous update. In asynchronous update, the procedures are executed as individual tasks and information is shared and used to guide the search for the optimal solution. The two traditional iteration strategies have their own strengths and weaknesses. The agents in synchronous update are able to consider the performance of the entire population before their next search step is determined. Therefore, the agents from synchronous update is stronger in exploitation, as the entire population is drawn towards a similar reference point, which is typically the population’s best performer. Meanwhile, an agent of asynchronous update is able to choose the reference point as soon as its fitness evaluation is finished. This update strategy improves the exploration of the population. Hence, selection of iteration strategy for a population-based metaheuristic can affect its overall performance. The aim of this study is to investigate the role and importance of iteration strategy towards population-based metaheuristics and to propose a new class of alternative iteration strategies that i) balances exploration and exploitation, and ii) avoid premature convergence without introducing extra complexity through combination of the traditional iteration strategies. Thus, a new class of iteration strategies which is a class of hybrid traditional strategies is proposed here. The strategies from this class are applicable for any population-based metaheuristics. The strategies are random switching, adaptive switching and adaptive switching with randomness. In the random switching strategy, the population randomly switches between the traditional strategies to cause disturbance to population diversity. The adaptive switching population, uses the information of the population’s condition to determine when to switch its iteration strategy. Meanwhile, the adaptive switching with randomness, embed randomness to encourage more number of switching. Experiments conducted using three parent algorithms namely particle swarm optimization (PSO), which is a popular population-based optimizer with population and individual memories, gravitational search algorithm (GSA), a memoryless young optimizer, and simulated Kalman filter (SKF), a newly introduced optimization algorithm that use population’s memory to guide an agent’s search, show that iteration strategy is an algorithm dependent parameter as well as function dependent. An iteration strategy is able to improve the performance of a parent algorithm and cause another parent algorithm to perform badly. The empirical analysis conducted here used the CEC2014’s benchmark functions for single objective optimization problems.

Actions (For repository staff only : Login required)