Fitness Estimation Strategy Assisted Competitive Swarm Optimizer for High Dimensional Expensive Problems
Chaoli Sun,Yaochu Jin,Jinliang Ding,Jianchao Zeng
DOI: https://doi.org/10.1145/2908961.2931713
2016-01-01
Abstract:Surrogate-assisted meta-heuristic algorithms are promising methods for solving computationally expensive optimization problems. It is usually assumed that the computational effort required for building and using surrogates is much less than that for calculating the fitness using the real fitness function. Different surrogate models, such as polynomials, kriging (also known as Gaussian process) and neural networks, have been proposed to assist population based meta-heuristics to solve computationally expensive problems [4][3][2]. In many cases, surrogates are meant for approximating the global fitness profile during the optimization, which however, may become infeasible, especially when the input dimension of the objective function, i.e., the number of decision variables, becomes high. Due to the limited number of training samples, global surrogates will become inaccurate and may introduce false optimums, which can mislead the evolutionary search. By contrast, local surrogate models aim to approximate a small region of the objective function, which is relatively easier to achieve with a limited number of training data. Although quite a large number of surrogate assisted meta-heuristic algorithms have been proposed, most ∗Corresponding author. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. GECCO’16 Companion, July 20–24, 2016, Denver, CO, USA. c ⃝ 2016 ACM. ISBN 978-1-4503-4323-7/16/07. . . $15.00 DOI: http://dx.doi.org/10.1145/2908961.2931713 of them work only for relatively low dimensional optimization problems. This might be attributed to the fact that it is extremely challenging to train an accurate surrogate model when the computational budget is limited for high dimensional problems. As far as we know, the maximum dimension of optimization problems solved by surrogate-assisted evolutionary algorithm is 50, which was reported in Liu et al. [5]. The authors proposed a Gaussian process surrogate model assisted differential evolution for medium-size computationally expensive optimization problems, in which principal component analysis was used to reduce the dimension. In this paper, we aim to develop a surrogate assisted evolutionary algorithm for solving large scale optimization problems having up to 500 decision variables. To this end, we adopted a recently proposed meta-heuristics, termed competitive swarm optimizer (CSO) [1] that was developed for large scale optimization. Since local surrogate models can more reliably approximate the fitness values than global ones, in this paper, we develop a new fitness inheritance technique, which can be seen a specific local surrogate model, to assist the CSO. Fitness inheritance was originally proposed for genetic algorithms [6] and was extended to PSO by Sun et al. [8], called FESPSO. In FESPSO, if the fitness of a particle is known, the fitness of its closest neighbour will be estimated according to the positional relationship between the two particles. Later on, a similarity measure was introduced into FESPSO in order to further reduce the number of expensive fitness evaluations [7]. Compared with computational model based surrogates, fitness estimation techniques proposed in [8] and [7] have the following advantages. First, these fitness estimation strategies do not require an external archive to save the historical information. Second, they are computationally much more efficient than most model based surrogate techniques. Finally, fit-