Analyzing the Role of 'Smart' Start Points in Coarse Search-Greedy Search

摘要

An inherent assumption in many search techniques is that information from existing solution(s) can help guide the search process to find better solutions. For example, memetic algorithms can use information from existing local optima to effectively explore a globally convex search space, and genetic algorithms assemble new solution candidates from existing solution components. At the extreme, the quality of a random solution may even be used to identify promising areas of the search space to explore. The best of several random solutions can be viewed as a "smart" start point for a greedy search technique, and the benefits of "smart" start points are demonstrated on several benchmark and real-world optimization problems. Although limitations exist, "smart" start points are most likely to be useful on continuous domain problems that have expensive solution evaluations.