Properties of Fitness Functions and Search Landscapes

Searching the Hyper-heuristic Design Space

“…Examples of offline learning for hyper-heuristics includes learning-classifier systems [17], case-based reasoning [18] and genetic programming [19]. While much work has been done in the offline analysis of landscapes and devising metrics of landscape quality ( [20], [21], [22], [23], [24], [25]), this information is not routinely used to inform online activity. As a simple but concrete example, the autocorrelation length of the landscape [26] could be used to inform parameters such as tabu-tenure or population size, but such interplay between design and experiment is not currently a routine part of metaheuristic development.…”

Heuristic algorithms and learning techniques: applications to the graph coloring problem

“…A research direction, especially in the context of evolutionary computing, consists of defining statistical measures to estimate the problem difficulty (i.e. convexity, ruggedness, smoothness or fitness distance correlation [Jones and Forrest, 1995]-see [Kallel et al, 2001;Merz, 2004] for a summary of such measures and related issues. Other studies deal with the structural similarities between local optima (i.e.…”

Study of Genetic Algorithms Behavior for High Epitasis and High Dimensional Deceptive Functions

“…The complexity of a problem is also obviously directly linked to the efficiency of the algorithms applied to solve it and to the capacity of these algorithms to exploit the properties of this problem (Jones, 1995). The first property influencing the complexity is the number of maxima of F (Kallel et al, 2001). For example, when |M| is large compared with |C|, an algorithm that evaluates points of C randomly and which has therefore an average complexity of |M|/|C|, is very efficient.…”