Dominance of meta-heuristics within a rapidly evolving world as of Knowledge--论文代写范文精选

在这篇essay代写范文中,我们审查解释MH的参数为什么仍不能令人满意。我们认为更引人注目的和全面的解释，实际上已经取得了成功,例如通过自定义一个特定问题的环境。下面的essay代写范文进行阐述。

Abstract
Although researchers often discuss the rising popularity of meta-heuristics (MH), there has been a paucity of data to directly support the notion that MH are growing in prominence compared to deterministic methods (DM). Here we provide the first evidence that MH usage is not only growing, but indeed appears to have surpassed DM as the algorithm framework of choice for solving optimization problems. Motivated by these findings, this paper aims to review and discuss the origins of meta-heuristic dominance. Explanations for meta-heuristic success are varied, however their robustness to variations in fitness landscape properties is often cited as an important advantage. In this paper, we review explanations for MH popularity and discuss why some of these arguments remain unsatisfying. 

We argue that a more compelling and comprehensive explanation would directly account for the manner in which most MH success has actually been achieved, e.g. through hybridization and customization to a particular problem environment. This paper puts forth the hypothesis that MH derive much of their utility from being flexible. This flexibility is empirically supported by evidence that MH design can adapt to a problem environment and can integrate domain knowledge. We propose what flexibility means from a search algorithm design context and we propose key attributes that should exist in a flexible algorithm framework. Interestingly, a number of these qualities are observed in robust biological systems. In light of these similarities, we consider whether the origins of biological robustness, (e.g. loose coupling, modularity, partial redundancy) could help to inspire the development of more flexible algorithm frameworks. We also discuss current trends in optimization problems and speculate that highly flexible algorithm frameworks will become increasingly popular within our diverse and rapidly changing world.

INTRODUCTION 
ATA on meta-heuristic usage in public, private, and academic sectors is sparse, however there has been some evidence that their use in computer-based problem solving is growing [1] [2] [3]. On almost a daily basis, there are new nature-inspired algorithms being proposed, new journals and conferences being advertised, as well as a continuous supply of new applications being considered within academic research. In [2], bibliographic data on genetic algorithms is used to show that publications within this field experienced a 40% annual growth from 1978 to 1998. More recently in [1], they present survey data showing that evolutionary computation (EC) usage is growing at a super linear rate in both public and private sectors. 

D Although these studies clearly indicate a growth in EC usage, it is not clear how these usage trends compare with similar research and development activity in deterministic methods. In particular, it has not been determined whether MH growth is actually outpacing alternative optimization techniques. By analyzing data from a number of publically accessible databases, we provide evidence in Box 1 that the usage of meta-heuristics is not only growing, but in many respects meta-heuristics are surpassing deterministic methods as the framework of choice for solving optimization problems. 

It is clear from the results in Box 1 that the number of optimization publications, case studies, and patents is growing and that this growth is in many ways irrespective of the search paradigm being considered. There are undoubtedly a number of interrelated factors contributing to this growth including technological innovation, global prosperity, as well as a growth in the number of problems that can be solved through computer-based methods, e.g. due to simulation technology and the growing availability of computing resources. However, it is also apparent from Box 1 that meta-heuristic implementation has been growing at a rate that is not matched by deterministic methods. 

Our aim in this paper is to try to understand why this is happening. Before concluding that MH’s newly found popularity is a reflection of utility, it is important to consider alternative explanations for this uneven growth. There are different reasons why a component, product or service grows in popularity within a competitive environment and not all of these are based on fitness. For instance, in evolving complex systems, the prominence (e.g. high connectivity) of a component within the system can sometimes be attributed to historical reasons. In particular, historical forces often bias growth in favour of past historical prominence, e.g. the well known “rich get richer” paradigm [4]. Comparing the rise in usage of the two optimization classes in Box 1, it is apparent that historical arguments can not account for the observed trends. Deterministic methods have a well-known rich history and were actively studied for decades prior to the first appearance of meta-heuristics. 

US patents of deterministic methods for solving linear programming and dynamic programming problems were first granted in 1972, while the first meta-heuristic (simulated annealing) was not patented until 1986. Taking the data from Figure 2a, for the ten years leading up to 1990 there were 2525 DM journal publications compared with only 208 for MH. Over the next ten years the relative size of this gap narrows (DM=15619, MH=8403), however the historical advantage at the turn of the century was still clearly in DM’s favour. Other plausible reasons for biased growth in favour of MH include superficial reasons such as the conceptual appeal of metaheuristics, e.g. the appeal of “nature-inspired” algorithms. 

Although this cannot be ruled out as a significant factor and may indeed account for some academic publication trends, conceptual appeal is less likely to explain trends in patents or the usage of MH in industry. In industry applications, those responsible for deciding which search techniques to implement should be primarily concerned with the anticipated efficacy of the algorithm framework and less concerned with any conceptual or otherwise superficial attachment. As we indicate in Box 2 for the specific case of genetic algorithms (GA) and industrial scheduling problems, there is considerable evidence that MH are being broadly implemented and that these implementations are often successful. Similar arguments might also apply to the patent trends shown in Figure 1. In this case, the costly decision to file for a patent is likely based on anticipated efficacy and not on superficial appeal. In short, the available data supports the conclusion that MH are being preferentially selected based on evidence of algorithm utility. 

However, this naturally raises the question of why MH are more useful for today’s problems. In the next section, we review fitness landscape-based arguments for understanding the utility of MH. We also review past arguments and evidence that MH success is strongly correlated with hybridization and customization efforts. In Section 3, we explore the hypothesis that the flexibility of an algorithm framework is the most important factor in determining the likelihood of algorithm success. We consider what algorithm flexibility means and the conditions that favor flexibility. This section also reviews trends taking place in industry and society and we speculate on important features to expect in future optimization techniques. In Section 4, we propose a theoretical basis for algorithm flexibility and discuss the relationship between these ideas and those developed in the study of complex adaptive systems. A summary of our main findings and arguments is given in Section 5 with experimental methods provided in Section 6.

EARLY EXPLANATIONS OF ALGORITHM UTILITY 
Early arguments in favour of MH focused on fitness landscape features or theories related to the operation of genetic algorithms, e.g. schema theory [5] and the building block hypothesis [6]. For instance, genetic algorithms were touted for their ability to deal with discontinuities in the fitness landscape, non-Gaussian noise in objective function measurements, nonstationarity of the fitness landscape, errors in determining objective function gradients, and numerical errors from computer calculations [7] [8] [9]. Their unabated success in multi-objective and multimodal fitness landscapes have also commonly been cited as important advantages. Furthermore, they often benefit from larger and more distributed computing resources; something that is increasingly available in both industry and academia. 

However, there is another narrative surrounding the success of MH that should be considered seriously when trying to understand the merits of these algorithms. As many successful algorithm developers repeatedly emphasize in conferences, workshops and lectures, an MH’s success or failure hinges upon the designer’s ability to integrate domain knowledge into the algorithm design and generally customize the algorithm to handle the particular needs of the customer and problem. This customization mantra extends beyond heuristic knowledge. The importance of customization to the success of a GA has been documented repeatedly over the last 15 years within reviews and individual studies . 

In the case of GA applied to industrial scheduling problems, which we review in Box 2, it is notable that almost all successful case studies involve a custom GA or GA hybrid. In the literature, it is common to find search operators that are custom-designed for a specific problem and that are effective at finding feasibility regions or more generally for finding useful parameter combinations in solution space. Domain knowledge is also frequently used to custom design stopping criteria, restart strategies, initialization strategies, constraint representation, and fitness functions, as well as to develop special encoding/decoding processes for solution representation. Acknowledging the influence that customization has had on the success of this field is important. In our previous statements, we implied that a specific set of MH algorithms are growing in popularity, which is not entirely accurate. 

A more accurate statement is that an ever diversifying set of algorithms labelled as MH are increasingly being used, and that many of these start off with a common algorithmic origin, e.g. the canonical genetic algorithm. Evidence of the importance of customization does not contradict previous claims that there are recognizable characteristics of problems that MH are particularly adept at handling. However it does suggest that the power of these problem characteristics to explain MH success is limited. Without accounting for the important role of algorithm design adaptation, fitness landscape arguments for algorithm success can have the unintended effect of implying that an MH is effective as an “off the shelf” or black-box optimization technique, so long as these problem characteristics are present. This broader statement is not at all supported by the available evidence.（essay代写)

51Due网站原创范文除特殊说明外一切图文著作权归51Due所有；未经51Due官方授权谢绝任何用途转载或刊发于媒体。如发生侵犯著作权现象，51Due保留一切法律追诉权。
更多essay代写范文欢迎访问我们主页 www.51due.com 当然有essay代写需求可以和我们24小时在线客服 QQ:800020041 联系交流。-X（essay代写)