A general cost-benefit-based adaptation framework for multimeme algorithms

Abstract

As Memetic Algorithms (MA) are a crossbreed between local searchers and Evolutionary Algorithms (EA) spreading of computational resources between evolutionary and local search is a key issue for a good performance, if not for success at all. This paper summarises and continues previous work on a general cost-benefit-based adaptation scheme for the choice of local searchers (memes), the frequency of their usage, and their search depth. This scheme eliminates the MA strategy parameters controlling meme usage, but raises new ones for steering the adaptation itself. Their impact is analysed and it will be shown that in the end the number of strategy parameters is decreased significantly as well as their range of meaningful values. In addition to this the number of fitness evaluations is reduced drastically. Both are necessary prerequisites for many practical applications as well as for the acceptance of the method by practitioners.Although the introduced framework is tailored to EAs producing more than one offspring per mating, it is also suited for those with only one child per pairing. So there are no preconditions to the EA for the described adaptation scheme to be applied. The benefit of LS within an MA can be summarised as follows: firstly, the already mentioned reduction of fitness evaluations required to achieve a certain quality and secondly, the possibility to include domain-specific knowledge into the search process. Additionally, offspring of poor performance, which are located in a region of attraction of a (local) optimum, are likely to be ignored in standard EAs, but can survive, if locally improved and support subsequent search. Furthermore, local search can work as a repair mechanism, if the genetic operators produce infeasible solutions in constraint optimisation and these solutions are penalised by reducing their fitness. The drawback consists in the following design questions, the last two of which mainly control the balance between global and local search or between exploration and exploitation: Keywords1. Which local searcher should be used?2. Should the LS result be used to update the genotype (Lamarckian evolution)or not (Baldwinian evolution)?3. How are offspring selected to undergo local search?4. How often should local improvement be applied or to which fraction of the generated offspring per generation? This is often referred to as local search frequency. [14,55,44]. When introducing the basic algorithms used for the experimental study, our answer to this question will be given. 3In this paper we present a cost-benefit-based adaptation scheme, which controls the parameters resulting from the remaining four of the above five questions. The beneficial effect of the adaptation scheme was reported about in [24,25], where also the new strategy parameters resulting from the scheme were introduced and discussed. In this work, the adaptation scheme is extended and the effect of the new strategy parameters is investigated empirically in more detail. The main goal is to find out their rele...

A self-adaptive Multimeme Memetic Algorithm co-evolving utility scores to control genetic operators and their parameter settings

“…Jakob [29] proposed a cost-benefit adaptation strategy for multi-meme algorithms in which adaptation is guided by the costs and benefits of a local search run. As a by-product, the strategy maintains a balance between intensification and diversification.…”

Fast Rescheduling of Multiple Workflows to Constrained Heterogeneous Resources Using Multi-Criteria Memetic Computing

“…The last difference concerns population structures and the concept of cellular MAs based on the diffusion model, for which two authors of [19] claim that they are -a new class of algorithms‖ [21]. Since the early 90's, our EA has been using a ring-shaped diffusion model, which is described in more detail in Section 4.2.1 [22][23][24]. This also holds for the memetic enhancements of our algorithm starting in 2001 [25].…”

“…The frequency of meme application is simply set to each pairing while the offspring selection can be steered by the parameter all_improvement to some extent. Both disadvantages mentioned above can be overcome by an adaptive parameter control and meme selection as described for example in [24,40,41]. Of course, good candidates for a set of memes must still be chosen manually and the adaptation figures out, which are better in which stage of the evolutionary process.…”

A self-adaptive memeplexes robust search scheme for solving stochastic demands vehicle routing problem

“…This workflow of memetic search proceeds until some stopping conditions are reached. Many of the current state-of-the-art MA are adaptive MAs that coordinate the memes while the evolutionary search progresses online, such as the meta-Lamarckian learning (Ong and Keane 2004), cost-benefit (Jakob 2010) or fitness diversity schemes (Caponio, Cascella, Neri, Salvatore, and Sumner 2007;Neri, Toivanen, Cascella, and Ong 2007;Caponio, Neri, and Tirronen 2009) and hyperheuristics (Segura, Miranda, and Leon 2011). Another class of adaptive MAs is the self-generated MAs and co-evolution MAs (Smith 2002;Krasnogor 2004), which rely on memetic evolution.…”