Adapting Genetic Algorithms for Combinatorial Optimization Problems in Dynamic Environments

Abstract

Combinatorial optimization problems (COPs) have a wide range of applications in engineering, operation research, and social sciences. Moreover, as real-time information and communication systems become increasingly available and the processing of real-time data becomes increasingly affordable, new versions of highly dynamic real-world applications are created. In such applications, information on the problem is not completely known a priori, but instead is revealed to the decision maker progressively with time. Consequently, solutions to different instances of a typical dynamic problem have to be found as time proceeds, concurrently with the incoming information. Given that the overwhelming majority of COPs are NP-hard, the presence of time and the associated uncertainty in their dynamic versions increases their complexity, making their dynamic versions even harder to solve than its static counterpart. However, environmental changes in real life typically do not alter the problem completely but affect only some part of the problem at a time. For example, not all vehicles break down at once, not all pre-made assignments are cancelled, weather changes affect only parts of roads, any other events like sickness of employees and machine breakdown do not happen concurrently. Thus, after an environmental change, there remains some information from the past that can be used for the future. Such problems call for a methodology to track their optimal solutions through time. The required algorithm should not only be capable of tackling combinatorial problems but should also be adaptive to changes in the environment. Evolutionary Algorithms (EAs) have been successfully applied to most COPs. Moreover, the ability of EAs to sample the search space, their ability to simultaneously manipulate a group of solutions, and their potential for adaptability increase their potential for dynamic problems. However, their tendency to converge prematurely in static problems and their lack of diversity in tracking optima that shift in dynamic environments are deficiencies that need to be addressed. Although many real world problems can be viewed as dynamic w e are interested only in those problems where the decision maker does not have prior knowledge of the complete problem, and hence the problem can not be so lved in advance. This article presents strategies to improve the ability of an algorithm to adapt to environmental changes, and more importantly to improve its efficiency at finding quality solutions. The first constructed