Exploratory Landscape Analysis of Continuous Space Optimization Problems Using Information Content

Algorithm selection for black-box continuous optimization problems: A survey on methods and challenges

“…ELA is an umbrella term for analytical, approximated and non-predictive methods [53,66,86,91,104] originally developed for combinatorial optimization problems [137]. For BCOPs, the existing methods are adaptations from their combinatorial counterparts [19,20,85,95,101,103,144,150], or purposely built for continuous spaces [19,83,91,98,120].…”

“…. ; y n f g , obtained by sorting a uniformly distributed sample [103]. The starting element in the sequence is a candidate from the sample selected at random, while the remaining elements in the sequence are selected using the nearest neighbor heuristic, i.e., the element whose Euclidean distance is the lowest to the current element.…”

“…However, this approach biases the sample unless n ! 1 [103]. A bias correction method, such as stratified or weighted sampling [26,101,170], is needed when a random walk with n ( 1 is used.…”

Analysing and comparing problem landscapes for black-box optimization via length scale

“…It is particularly useful in applications where the cost of evaluating f is high, as it ensures a wide variety of parameter value combinations, at a reasonable computation cost. In the context of landscape analysis, both Mersmann et al [113] and Muñoz et al [118] have recently utilised LH sampling to calculate various landscape features of BBOB problems spanning between 2 to 20 dimensions.…”

“…Dispersion has been used to study the performance of algorithms relative to particular problems (and their structure); CMA-ES, hybrid PSO/CMA-ES algorithms, Pattern Search methods and Local Search have been analysed on a number of benchmark functions [103,121,202]. The dispersion values of 2, 5, 10 and 20 dimensional problems from the BBOB problem set have also been used in the feature-set of an algorithm prediction model [118,119]. Garden …”