DOCSLIB.ORG

Reconstructing Damaged Complex Networks Based on Neural Networks

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Reconstructing Damaged Complex Networks Based on Neural Networks S S symmetry Article Reconstructing Damaged Complex Networks Based on Neural Networks Ye Hoon Lee 1 and Insoo Sohn 2,* 1 Department of Electronic and IT Media Engineering, Seoul National University of Science and Technology, Seoul 01811, Korea; [email protected] 2 Division of Electronics and Electrical Engineering, Dongguk University, Seoul 04620, Korea * Correspondence: [email protected]; Tel.: +82-2-2260-8604 Received: 2 October 2017; Accepted: 9 December 2017; Published: 9 December 2017 Abstract: Despite recent progress in the study of complex systems, reconstruction of damaged networks due to random and targeted attack has not been addressed before. In this paper, we formulate the network reconstruction problem as an identification of network structure based on much reduced link information. Furthermore, a novel method based on multilayer perceptron neural network is proposed as a solution to the problem of network reconstruction. Based on simulation results, it was demonstrated that the proposed scheme achieves very high reconstruction accuracy in small-world network model and a robust performance in scale-free network model. Keywords: network reconstruction; neural networks; small world networks; scale free networks 1. Introduction Complex networks have received growing interest from various disciplines to model and study the network topology and interaction between nodes within a modeled network [1–3]. One of the important problems that are actively studied in the area of network science is the robustness of a network under random failure of nodes and intentional attack on the network. For example, it has been found that scale-free networks are more robust compared to random networks against random removal of nodes, but are more sensitive to targeted attacks [4–6]. Furthermore, many approaches have been proposed to optimize conventional networks against random failure and intentional attack compared to the conventional complex networks [7–10]. However, these approaches have been mainly concentrated on designing network topology based on various optimization techniques to minimize the damage to the network. So far, no work has been reported on techniques to repair and recover the network topology after the network has been damaged. Numerous solutions have been proposed on reconstruction of spreading networks based on spreading data [11–13], but the problem on the reconstruction of damaged networks due to random and targeted attacks has not been addressed before. Artificial neural networks (NNs) have been applied to solve various problems in complex systems due to powerful generalization abilities of NNs [14]. Some of the applications where NNs have been successfully used are radar waveform recognition [15], image recognition [16,17], indoor localization [18,19], and peak-to-average power reduction [20,21]. In this paper, we propose a novel network reconstruction method based on the NN technique. To the best of our knowledge, this work is the first attempt to recover the network topology after the network has been damaged and also the first attempt to apply NN technique for complex network optimization. We formulate the network reconstruction problem as an identification of a network structure based on a much reduced amount of link information contained in the adjacency matrix of the damaged network. The problem is especially challenging due to (1) very large number of possible network configurations 2 on the order of 2N , where N is the number of nodes and (2) very small number of node interaction Symmetry 2017, 9, 310; doi:10.3390/sym9120310 www.mdpi.com/journal/symmetry Symmetry 2017, 9, 310 2 of 11 Symmetryespecially2017 challenging, 9, 310 due to (1) very large number of possible network configurations on the 2order of 11 2 of 2N , where N is the number of nodes and (2) very small number of node interaction information informationdue to node due removals. to node We removals. simplify We the simplify problem the by problem the following by the assumptions following assumptions (1) average (1) number average of numberreal connections of real connections of a network of is a networkmuch smaller is much than smaller all thethan possible all the link possible configurations link configurations and (2) link andinformation (2) link information of M undamaged of M undamaged networks are networks available. are available.Based on these Based assumptions, on these assumptions, we chose wethe chosemultiple the-layer multiple-layer perceptron perceptron neural network neural network (MLPNN), (MLPNN), which whichis one isof one the of frequently the frequently used used NN NNtechniques, techniques, as the as basis the basis of our of method. our method. We evaluate We evaluate the performance the performance of the of proposed the proposed method method based basedon simulations on simulations in two in twoclassical classical complex complex networks networks (1) (1) small small-world-world network network and and (2) scale-freescale-free networks,networks, generatedgenerated by by Watts Watts and and Strogatz Strogatz model model and and Barab Barabásiási and and Albert Albert model, model, respectively. respectively. TheThe restrest of the the paper paper is is organized organized as asfollows. follows. Section Section 2 describes2 describes the small the small-world-world network network model modeland scale and-free scale-free network, network, followed followed by the by thenetwork network damage damage model. model. In In Section Section 3,3, wewe proposepropose the the reconstructionreconstruction methodmethod basedbased onon NNNN technique.technique. InIn SectionSection4 ,4, we we present present the the numerical numerical results results of of the the proposedproposed method, method, and and conclusions conclusions are are given given Section Section5 .5. 2.2. ModelModel 2.1.2.1. Small-WorldSmall-World NetworkNetwork Small-worldSmall-world networknetwork isis anan importantimportant networknetwork modelmodel withwith lowlow averageaverage pathpath lengthlength andand highhigh clusteringclustering coefficientcoefficient [22[22––2424].]. Small-worldSmall-world networksnetworks havehave homogeneoushomogeneous networknetwork topologytopology withwith aa degreedegree distribution distribution approximated approximated by by the the Poisson Poisson distribution. distribution. A small-world A small-worl networkd network is created is created based onbased a regular on a regular lattice such lattice a ring such of aN ringnodes, of N where nodes, each where node each is connected node is connected to J nearest to nodes. J nearest Next, nodes. the linksNext, are the randomly links are rewired randomly to one rewired of N nodes to one in theof N network nodes within the probability network withp. The probability rewiring process p. The isrewiring repeated process for all is the repeated nodes nfor= 1all... theN nodes. By controlling n = 1 … N. theBy controlling rewiring probability the rewiringp, the probability network p will, the interpolatenetwork will between interpolate a regular between lattice a regular (p = 0) tolattice a random (p = 0) network to a random (p = 1).network Figure (1p shows= 1). Figure the detailed 1 shows algorithmthe detailed for algorithm small-world for networksmall-world construction network inconstruction pseudocode in format.pseudocode format. 1: Initialization: 2: Set the total number of nodes N 3: Set the rewiring probability p 4: Set the node degree K 5: end initialization 6: 7: for each node n (1, N) 8: Connect to nearest K nodes. 9: end for 10: 11: for each node n (1, N) 12: for each link k (1, K) 13: Generate a random number r (0, 1) 14: if r < p 15: Select a node randomly m (1, N) 16: Rewire to node m 17: end if 18: end for 19: end for FigureFigure1. 1.Small-world Small-world networknetwork algorithm.algorithm. 2.2.2.2. Scale-FreeScale-Free NetworkNetwork Scale-freeScale-free networks, networks, such such as asBarab Barabáásisi and and AlbertAlbert (BA)(BA) network,network, areare evolvingevolving networksnetworks thatthat havehave degreedegree distributiondistribution followingfollowing power-lawpower-lawmodel model[ 4[4––66].]. Scale-freeScale-free networksnetworks consistconsist ofof aa smallsmall numbernumber ofof nodesnodes withwith veryvery high degree of of connections connections and and rest rest of of the the nodes nodes with with low low degree degree connections. connections. A Ascale scale-free-free network network starts starts the the evolution evolution process process with a small number ofof mm00 nodes. Next, aa newnew nodenode isis introduced to the network and attaches to m existing nodes with high degree k. The process consisting Symmetry 2017, 9, 310 3 of 11 Symmetry 2017, 9, 310 3 of 11 introduced to the network and attaches to m existing nodes with high degree k. The process consisting of new node introduction and preferential attachment is repeated until a network with N = t + m0 of new node introduction and preferential attachment is repeated until a network with N = t + m0 nodesnodes hashas been been constructed. constructed. Figure Figure2 shows 2 theshows detailed the algorithmdetailed foralgorithm scale-free for network scale-free construction network inconstruction pseudocode in format. pseudocode format. 1: Initialization: 2: Set the initial number of nodes m0 3: Set the total number of nodes N 4: Set the node degree K 5: end initialization 6: 7: for each node n (1, m0) 8: Connect to nearest K nodes. 9: end for 10: 11: for each new node 12: Calculate the
Load more

Recommended publications
  • Networkx: Network Analysis with Python
    NetworkX: Network Analysis with Python Salvatore Scellato Full tutorial presented at the XXX SunBelt Conference “NetworkX introduction: Hacking social networks using the Python programming language” by Aric Hagberg & Drew Conway Outline 1. Introduction to NetworkX 2. Getting started with Python and NetworkX 3. Basic network analysis 4. Writing your own code 5. You are ready for your project! 1. Introduction to NetworkX. Introduction to NetworkX - network analysis Vast amounts of network data are being generated and collected • Sociology: web pages, mobile phones, social networks • Technology: Internet routers, vehicular flows, power grids How can we analyze this networks? Introduction to NetworkX - Python awesomeness Introduction to NetworkX “Python package for the creation, manipulation and study of the structure, dynamics and functions of complex networks.” • Data structures for representing many types of networks, or graphs • Nodes can be any (hashable) Python object, edges can contain arbitrary data • Flexibility ideal for representing networks found in many different fields • Easy to install on multiple platforms • Online up-to-date documentation • First public release in April 2005 Introduction to NetworkX - design requirements • Tool to study the structure and dynamics of social, biological, and infrastructure networks • Ease-of-use and rapid development in a collaborative, multidisciplinary environment • Easy to learn, easy to teach • Open-source tool base that can easily grow in a multidisciplinary environment with non-expert users
    [Show full text]
  • Emerging Topics in Internet Technology: a Complex Networks Approach
    1 Emerging Topics in Internet Technology: A Complex Networks Approach Bisma S. Khan 1, Muaz A. Niazi *,2 COMSATS Institute of Information Technology, Islamabad, Pakistan Abstract —Communication networks, in general, and internet technology, in particular, is a fast-evolving area of research. While it is important to keep track of emerging trends in this domain, it is such a fast-growing area that it can be very difficult to keep track of literature. The problem is compounded by the fast-growing number of citation databases. While other databases are gradually indexing a large set of reliable content, currently the Web of Science represents one of the most highly valued databases. Research indexed in this database is known to highlight key advancements in any domain. In this paper, we present a Complex Network-based analytical approach to analyze recent data from the Web of Science in communication networks. Taking bibliographic records from the recent period of 2014 to 2017 , we model and analyze complex scientometric networks. Using bibliometric coupling applied over complex citation data we present answers to co-citation patterns of documents, co-occurrence patterns of terms, as well as the most influential articles, among others, We also present key pivot points and intellectual turning points. Complex network analysis of the data demonstrates a considerably high level of interest in two key clusters labeled descriptively as “social networks” and “computer networks”. In addition, key themes in highly cited literature were clearly identified as “communication networks,” “social networks,” and “complex networks”. Index Terms —CiteSpace, Communication Networks, Network Science, Complex Networks, Visualization, Data science.
    [Show full text]
  • Social Network Analysis and Information Propagation: a Case Study Using Flickr and Youtube Networks
    International Journal of Future Computer and Communication, Vol. 2, No. 3, June 2013 Social Network Analysis and Information Propagation: A Case Study Using Flickr and YouTube Networks Samir Akrouf, Laifa Meriem, Belayadi Yahia, and Mouhoub Nasser Eddine, Member, IACSIT 1 makes decisions based on what other people do because their Abstract—Social media and Social Network Analysis (SNA) decisions may reflect information that they have and he or acquired a huge popularity and represent one of the most she does not. This concept is called ″herding″ or ″information important social and computer science phenomena of recent cascades″. Therefore, analyzing the flow of information on years. One of the most studied problems in this research area is influence and information propagation. The aim of this paper is social media and predicting users’ influence in a network to analyze the information diffusion process and predict the became so important to make various kinds of advantages influence (represented by the rate of infected nodes at the end of and decisions. In [2]-[3], the marketing strategies were the diffusion process) of an initial set of nodes in two networks: enhanced with a word-of-mouth approach using probabilistic Flickr user’s contacts and YouTube videos users commenting models of interactions to choose the best viral marketing plan. these videos. These networks are dissimilar in their structure Some other researchers focused on information diffusion in (size, type, diameter, density, components), and the type of the relationships (explicit relationship represented by the contacts certain special cases. Given an example, the study of Sadikov links, and implicit relationship created by commenting on et.al [6], where they addressed the problem of missing data in videos), they are extracted using NodeXL tool.
    [Show full text]
  • Complex Networks in Computer Vision
    Complex networks in computer vision Pavel Voronin Shape to Network [Backes 2009], [Backes 2010a], [Backes 2010b] Weights + thresholds Feature vectors Features = characteris>cs of thresholded undirected binary networks: degree or joint degree (avg / min / max), avg path length, clustering coefficient, etc. Mul>scale Fractal Dimension Properes • Only uses distances => rotaon invariant • Weights normalized by max distance => scale invariant • Uses pixels not curve elements => robust to noise and outliers => applicable to skeletons, mul>ple contours Face recogni>on [Goncalves 2010], [Tang 2012a] Mul>ple binarizaon thresholds Texture to Network [Chalumeau 2008], [Backes 2010c], [Backes 2013] Thresholds + features Features: degree, hierarchical degree (avg / min / max) Invariance, robustness Graph structure analysis [Tang 2012b] Graph to Network Thresholds + descriptors Degree, joint degree, clustering-distance Saliency Saccade eye movements + different fixaon >me = saliency map => model them as walks in networks [Harel 2006], [Costa 2007], [Gopalakrishnan 2010], [Pal 2010], [Kim 2013] Network construc>on Nodes • pixels • segments • blocks Edges • local (neighbourhood) • global (most similar) • both Edge weights = (distance in feature space) / (distance in image space) Features: • relave intensity • entropy of local orientaons • compactness of local colour Random walks Eigenvector centrality == staonary distribu>on for Markov chain == expected >me a walker spends in the node Random walk with restart (RWR) If we have prior info on node importance,
    [Show full text]
  • Analysis of Statistical and Structural Properties of Complex Networks with Random Networks
    Appl. Math. Inf. Sci. 11, No. 1, 137-146 (2017) 137 Applied Mathematics & Information Sciences An International Journal http://dx.doi.org/10.18576/amis/110116 Analysis of Statistical and Structural Properties of Complex networks with Random Networks S. Vairachilai1,∗, M. K. Kavitha Devi2 and M. Raja 1 1 Department of CSE, Faculty of Science and Technology, IFHE Hyderabad, Telangana -501203, India. 2 Department of CSE, ThiagarajarCollege of Engineering, Madurai,Tamil Nadu-625015, India. Received: 2 Sep. 2016, Revised: 25 Oct. 2016, Accepted: 29 Oct. 2016 Published online: 1 Jan. 2017 Abstract: Random graphs are extensive, in addition, it is used in several functional areas of research, particularly in the field of complex networks. The study of complex networks is a useful and active research areas in science, such as electrical power grids and telecommunication networks, collaboration and citation networks of scientists,protein interaction networks, World-Wide Web and Internet Social networks, etc. A social network is a graph in which n vertices and m edges are selected at random, the vertices represent people and the edges represent relationships between them. In network analysis, the number of properties is defined and studied in the literature to identify the important vertex in a network. Recent studies have focused on statistical and structural properties such as diameter, small world effect, clustering coefficient, centrality measure, modularity, community structure in social networks like Facebook, YouTube, Twitter, etc. In this paper, we first provide a brief introduction to the complex network properties. We then discuss the complex network properties with values expected for random graphs.
    [Show full text]
  • Arxiv:2006.02870V1 [Cs.SI] 4 Jun 2020
    The why, how, and when of representations for complex systems Leo Torres Ann S. Blevins [email protected] [email protected] Network Science Institute, Department of Bioengineering, Northeastern University University of Pennsylvania Danielle S. Bassett Tina Eliassi-Rad [email protected] [email protected] Department of Bioengineering, Network Science Institute and University of Pennsylvania Khoury College of Computer Sciences, Northeastern University June 5, 2020 arXiv:2006.02870v1 [cs.SI] 4 Jun 2020 1 Contents 1 Introduction 4 1.1 Definitions . .5 2 Dependencies by the system, for the system 6 2.1 Subset dependencies . .7 2.2 Temporal dependencies . .8 2.3 Spatial dependencies . 10 2.4 External sources of dependencies . 11 3 Formal representations of complex systems 12 3.1 Graphs . 13 3.2 Simplicial Complexes . 13 3.3 Hypergraphs . 15 3.4 Variations . 15 3.5 Encoding system dependencies . 18 4 Mathematical relationships between formalisms 21 5 Methods suitable for each representation 24 5.1 Methods for graphs . 24 5.2 Methods for simplicial complexes . 25 5.3 Methods for hypergraphs . 27 5.4 Methods and dependencies . 28 6 Examples 29 6.1 Coauthorship . 29 6.2 Email communications . 32 7 Applications 35 8 Discussion and Conclusion 36 9 Acknowledgments 38 10 Citation diversity statement 38 2 Abstract Complex systems thinking is applied to a wide variety of domains, from neuroscience to computer science and economics. The wide variety of implementations has resulted in two key challenges: the progenation of many domain-specific strategies that are seldom revisited or questioned, and the siloing of ideas within a domain due to inconsistency of complex systems language.
    [Show full text]
  • Deep Generative Modeling in Network Science with Applications to Public Policy Research
    Working Paper Deep Generative Modeling in Network Science with Applications to Public Policy Research Gavin S. Hartnett, Raffaele Vardavas, Lawrence Baker, Michael Chaykowsky, C. Ben Gibson, Federico Girosi, David Kennedy, and Osonde Osoba RAND Health Care WR-A843-1 September 2020 RAND working papers are intended to share researchers’ latest findings and to solicit informal peer review. They have been approved for circulation by RAND Health Care but have not been formally edted. Unless otherwise indicated, working papers can be quoted and cited without permission of the author, provided the source is clearly referred to as a working paper. RAND’s R publications do not necessarily reflect the opinions of its research clients and sponsors. ® is a registered trademark. CORPORATION For more information on this publication, visit www.rand.org/pubs/working_papers/WRA843-1.html Published by the RAND Corporation, Santa Monica, Calif. © Copyright 2020 RAND Corporation R® is a registered trademark Limited Print and Electronic Distribution Rights This document and trademark(s) contained herein are protected by law. This representation of RAND intellectual property is provided for noncommercial use only. Unauthorized posting of this publication online is prohibited. Permission is given to duplicate this document for personal use only, as long as it is unaltered and complete. Permission is required from RAND to reproduce, or reuse in another form, any of its research documents for commercial use. For information on reprint and linking permissions, please visit www.rand.org/pubs/permissions.html. The RAND Corporation is a research organization that develops solutions to public policy challenges to help make communities throughout the world safer and more secure, healthier and more prosperous.
    [Show full text]
  • Complex Design Networks: Structure and Dynamics
    Complex Design Networks: Structure and Dynamics Dan Braha New England Complex Systems Institute, Cambridge, MA 02138, USA University of Massachusetts, Dartmouth, MA 02747, USA Email: [email protected] Why was the $6 billion FAA air traffic control project scrapped? How could the 1977 New York City blackout occur? Why do large scale engineering systems or technology projects fail? How do engineering changes and errors propagate, and how is that related to epidemics and earthquakes? In this paper we demonstrate how the rapidly expanding science of complex design networks could provide answers to these intriguing questions. We review key concepts, focusing on non-trivial topological features that often occur in real-world large-scale product design and development networks; and the remarkable interplay between these structural features and the dynamics of design rework and errors, network robustness and resilience, and design leverage via effective resource allocation. We anticipate that the empirical and theoretical insights gained by modeling real-world large-scale product design and development systems as self-organizing complex networks will turn out to be the standard framework of a genuine science of design. 1. Introduction: the road to complex networks in engineering design Large-scale product design and development is often a distributed process, which involves an intricate set of interconnected tasks carried out in an iterative manner by hundreds of designers (see Fig. 1), and is fundamental to the creation of complex man- made systems [1—16]. This complex network of interactions and coupling is at the heart of large-scale project failures as well as of large-scale engineering and software system failures (see Table 1).
    [Show full text]
  • Random Graph Models of Social Networks
    Colloquium Random graph models of social networks M. E. J. Newman*†, D. J. Watts‡, and S. H. Strogatz§ *Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501; ‡Department of Sociology, Columbia University, 1180 Amsterdam Avenue, New York, NY 10027; and §Department of Theoretical and Applied Mechanics, Cornell University, Ithaca, NY 14853-1503 We describe some new exactly solvable models of the structure of actually connected by a very short chain of intermediate social networks, based on random graphs with arbitrary degree acquaintances. He found this chain to be of typical length of distributions. We give models both for simple unipartite networks, only about six, a result which has passed into folklore by means such as acquaintance networks, and bipartite networks, such as of John Guare’s 1990 play Six Degrees of Separation (10). It has affiliation networks. We compare the predictions of our models to since been shown that many networks have a similar small- data for a number of real-world social networks and find that in world property (11–14). some cases, the models are in remarkable agreement with the data, It is worth noting that the phrase ‘‘small world’’ has been used whereas in others the agreement is poorer, perhaps indicating the to mean a number of different things. Early on, sociologists used presence of additional social structure in the network that is not the phrase both in the conversational sense of two strangers who captured by the random graph. discover that they have a mutual friend—i.e., that they are separated by a path of length two—and to refer to any short path social network is a set of people or groups of people, between individuals (8, 9).
    [Show full text]
  • Navigating Networks by Using Homophily and Degree
    Navigating networks by using homophily and degree O¨ zgu¨r S¸ ims¸ek* and David Jensen Department of Computer Science, University of Massachusetts, Amherst, MA 01003 Edited by Peter J. Bickel, University of California, Berkeley, CA, and approved July 11, 2008 (received for review January 16, 2008) Many large distributed systems can be characterized as networks We show that, across a wide range of synthetic and real-world where short paths exist between nearly every pair of nodes. These networks, EVN performs as well as or better than the best include social, biological, communication, and distribution net- previous algorithms. More importantly, in the majority of cases works, which often display power-law or small-world structure. A where previous algorithms do not perform well, EVN synthesizes central challenge of distributed systems is directing messages to whatever homophily and degree information can be exploited to specific nodes through a sequence of decisions made by individual identify much shorter paths than any previous method. nodes without global knowledge of the network. We present a These results have implications for understanding the func- probabilistic analysis of this navigation problem that produces a tioning of current and prospective distributed systems that route surprisingly simple and effective method for directing messages. messages by using local information. These systems include This method requires calculating only the product of the two social networks routing messages, referral systems for finding measures widely used to summarize all local information. It out- informed experts, and also technological systems for routing performs prior approaches reported in the literature by a large messages on the Internet, ad hoc wireless networking, and margin, and it provides a formal model that may describe how peer-to-peer file sharing.
    [Show full text]
  • Scale- Free Networks in Cell Biology
    Scale- free networks in cell biology Réka Albert, Department of Physics and Huck Institutes of the Life Sciences, Pennsylvania State University Summary A cell’s behavior is a consequence of the complex interactions between its numerous constituents, such as DNA, RNA, proteins and small molecules. Cells use signaling pathways and regulatory mechanisms to coordinate multiple processes, allowing them to respond to and adapt to an ever-changing environment. The large number of components, the degree of interconnectivity and the complex control of cellular networks are becoming evident in the integrated genomic and proteomic analyses that are emerging. It is increasingly recognized that the understanding of properties that arise from whole-cell function require integrated, theoretical descriptions of the relationships between different cellular components. Recent theoretical advances allow us to describe cellular network structure with graph concepts, and have revealed organizational features shared with numerous non-biological networks. How do we quantitatively describe a network of hundreds or thousands of interacting components? Does the observed topology of cellular networks give us clues about their evolution? How does cellular networks’ organization influence their function and dynamical responses? This article will review the recent advances in addressing these questions. Introduction Genes and gene products interact on several level. At the genomic level, transcription factors can activate or inhibit the transcription of genes to give mRNAs. Since these transcription factors are themselves products of genes, the ultimate effect is that genes regulate each other's expression as part of gene regulatory networks. Similarly, proteins can participate in diverse post-translational interactions that lead to modified protein functions or to formation of protein complexes that have new roles; the totality of these processes is called a protein-protein interaction network.
    [Show full text]
  • Optimizing the Topology of Complex Neural Networks
    Optimising the topology of complex neural networks Fei Jiang1;2, Hugues Berry1, and Marc Schoenauer2 1Project-team Alchemy and 2Project-team TAO, INRIA Futurs, Parc Club Orsay Universite, 91893 Orsay Cedex France, [email protected] Summary. In this paper, we study instances of complex neural networks, i.e. neural networks with complex topologies. We use Self-Organizing Map neural networks whose neighbourhood relationships are defined by a complex network, to classify handwritten digits. We show that topology has a small impact on performance and robustness to neuron failures, at least at long learning times. Performance may however be increased (by almost 10%) by artificial evolution of the network topology. In our experimental conditions, the evolved networks are more random than their parents, but display a more heterogeneous degree distribution. 1 Introduction The connectivity structure of complex networks (i.e. their topology) is a cru- cial determinant of information transfer in large networks (internet, social networks, metabolic networks...) [1]. Hence, the computation made by com- plex neural networks, i.e. neural networks with complex connectivity struc- ture, could be dependent on their topology. For instance, recent studies have shown that introducing a small-world topology in a multilayer perceptron increases its performance [13, 2]. However, other studies have inspected the performance of Hopfield [5, 10, 9, 14] or Echo state networks [3] with small- world or scale-free topologies and reported more contrasted results. Using artificial evolutionary algorithms to modify the topology of neu- ral networks so as to optimise their performance has become widespread in the artificial neural networks community for several years [11, 16].
    [Show full text]