Frequency Fitness Assignment

摘要: Metaheuristic optimization procedures such as evolutionary algorithms are usually driven by an objective function that rates the quality of a candidate solution. However, it is not clear in practice whether adequately rewards intermediate solutions on path to global optimum and may exhibit deceptiveness, epistasis, neutrality, ruggedness, lack causality. In this paper, we introduce frequency fitness H, subject minimization, which how often with same value have been discovered so far. The ideas behind method good difficult find if algorithm gets stuck at local optimum, values surrounding will increase over time, eventually allow leave region again. We substitute assignment process (FFA) for into several different algorithms. conduct comprehensive set experiments: synthesis genetic programming (GP), solution MAX-3SAT problems algorithms, classification Memetic Genetic Programming, numerical (1+1) Evolution Strategy, verify utility FFA. Given they no access original all, surprising some (e.g., task) FFA-based variants perform significantly better. cannot be guaranteed all tested problems. Thus, also analyze scenarios where using FFA do better or even worse than functions.