Using Evolutionary Algorithms to Solve Problems by Combining Choices of Heuristics

Abstract

Potential users of evolutionary algorithms (EAs) are often deterred by the ‘black box’ nature of many of the available examples. Typically an evolutionary algorithm is designed to search a problem’s solution space directly, and the user simply waits for some stopping criterion to take effect. However, the user usually gets no guarantees about the quality of the fittest solution at that point. It is unsurprising, therefore, that users may choose to use some simpler, cheaper heuristic method whose performance is better understood and faster even though it may well deliver poorer results. Commercial users in particular often have good reason to be nervous about using EAs in situations in which their business is likely to be judged on the quality of the EA’s result. However, there are ways in which they can still use EAs to very good effect. This paper discusses one such way, namely using an EA to choose which heuristics to apply at each stage in some sequential decision process. If the available heuristics are individually acceptable, then a combination of them is going to produce a better quality result than any of them individually would. This can be guaranteed by the simple device of seeding the initial population with chromosomes that employ just one heuristic throughout. The paper describes some examples of this approach and discusses possible developments of the idea.