Vous êtes sur la page 1sur 32

Taylor & Francis

Internalional Journa/ of Production Research. Vol. 43, No. 20. 15 October 2005, 4235-4265

K^

Supply-chain networks: a complex adaptive systems perspeetive AMIT SURANAt. SOUNDAR KUMARA*t MARK GREAVES and USHA NANDINI RAGHAVAN|
tDeparimetit of Mi;t;hiiiiical Engineering. The Massachusetts Institute of Technology. Cambridge. MA 02139. LJSA J3I() Leonhiird Building. The Hiirold ;ind Inge Marcus Department of Industrial & Manufacturing Engineering, The Pennsylvania State University. University Park, PA 16802. USA SIXO, DARPA. 3701 North Fairfax Drive. Arlinglon, VA 22203-1714. USA

(Revision received May 2005) In lliis era. information technology is revolutionizing almost every domain of technology and society, whereas the 'complexity revolution* is occurring in science at a silent pace. In this paper, we look at Ihe impact of the two. in the context oi' supply-chain networks. With the advent of information technology, supply chains have acquired a complexity almost equivalent to that of biological systems. However, one of the major challenges that we are facing in supply-chain management is the deployment of cooaiination strategies that lead to adaptive, flexible and coherent collective behaviour in supply chains. The main hurdle has been the lack of the principles that govern how supply chains with complex organizational structure and function arise and develop, and what organizations and functionality are attainable, given specific kinds of lower-level constituent entities. The study of Complex Adaptive Systems (CAS), has been a research effort attempting to find common characteristics and/or formal distinctions among cotnplex systems arising in diverse domains (like biology, social systems, ecology and technology) that might lead to a better understanding of how complexity occurs, whether it follows any general seientific laws of nature, and how it might be related lo simplicity. In this paper, we argue that supply chains should be treated as a CAS. With this recognition, we propose how various concepts, tools and techniques used in the study of CAS ean be exploited to characterize and tnodel supply-chain networks. These tools and techniques are based on the fields of nonlinear dynamics, statistical physics and information theor>, Kevword.s: Supply chain; Complexity; Complex adaptive systems; Nonlinear dynamics; Networks

1. Introduction A supply chain i.s a complex network with an overwhelming number of interactions and inter-depetidencies among different entities, processes and resources. The network is highly nonlinear, shows complex multi-scale behaviour, has a structure spatiniiig several scales, and evolves and self-organizes through u complex interplay

*Corresponding author. Email: skumara( psu.cdu


liiti-riuilioiml Jiiiiriiiil n/ Pnicliittioii Ri'.wanh ISSN 0020 7543 prim ISSN 1.166 S8KX ontine r 20(15 T;iyli>r & Hrancis http:. .www.tiindf.co uk Journals DOI: 10. tO8O 00207540500142274

4236

A. Surana et al.

of its structure and function. This sheer complexity of supply-chain networks, with inevitable laek of prediction, makes it diliicult to manage and eontrol (hem. Furthermore, the changing organizational and market trends mean that the supply ehains should be highly dynamic, scalable, reconfigurable. agile and adaptive: the network should sense and respond eifectively and efficiently to satisfy customer demand. Supply-chain management necessitates the deeisions made by business entities to consider more factors that are global. The sueeessful integration of the entire supply-chain process now depends heavily on the availability of accurate and timely information thai can be shared by all members of the supply chain. Information technology, with its capability of setting up dynamic information exchange networks, has been a key enabling factor in shaping supply chains to meet such requirements. However, a major obstacle remains in the deployment of coordination and decision technologies to achieve complex, adaptive, and flexible collective behaviour in the network. This is due to the lack of our understanding of organizational, functional and evolutionary aspects in supply chains. A key realization to tackle this problem is that supply-chain networks should be treated not just as a 'system' but as a 'Complex Adaptive System" (CAS). The study of CAS augments the systems theory and provides a rich set of tools and techniques to model and analyse the complexity arising in systems encompassing science and technology. In this paper, we take this perspective in dealing with supply ehains and show how various advances in the realm of CAS provide novel and ettective ways to characterize, understand and manage their emergent dynamics, A similar viewpoint has been emphasized by Choi cl al. (2001). who aimed lo demonstrate how supply networks should be managed if we recognize them as CAS. The eoncept of CAS allows one to understand how supply networks as living systems co-evolve with the rugged and dynamic environment in which they exist and identify patterns that arise in such an evolution. The authors conjecture various propositions stating how the patterns of behaviour of individual agents in a supply network relate to the emergent dynamic's of the network. One of the important deductions made is that when managing supply networks, managers must appropriately balance how much to control, and how much to let emerge. However, no concrete framework has been suggested under which such conjectures can be verified and generalized. It is the goal of this paper to show how the theoretical advances made in the realm of CAS can be used to study such issues systematically and formally in the context of supply-ehain networks. We posit that supply chains are complex adaptive systems. However, we do not provide conclusive proofs for such a claim. We survey the emerging literature, faithfully report on the state of the art in CAS and try to establish connections, as much as possible, between CAS tools and supply-ehain analysis. Through our effort, we would like to pave research directions in supply ehains from a CAS point of view. This paper is divided into eight sections. In seetion 2. we give a brief introduction to complex adaptive systems in which we discuss the arehitecture and characteristics of complex sy.stems in diverse areas encompassing biology, social systems, ecology and technology. In section 3, we discuss characteristics of supply chain-networks and argue that they should be understood in terms of a CAS. We also present some emerging trends in supply chains and the increasing critical role of information technology in supply-ehain management in tbe ligbt of these trends. In seetion 4, we give a brief overview of the main techniques used for modellmg and analysis

Supplv-ihain nelnork.s: a comple.x adaptive systems perspective

4237

of supply chains and then discuss how the science of complexity provides a genuine extension and reformulation oi' these appmaches. Like any CAS, the study ot supply chains should involve a proper balance of simulation and theory. System dynamics-based and recently agent-based simulation models (inspired by complexity theory) are used extensively lo make theoretical investigations of supply chains feasible and to support decision-making in real-world supply chains. A system dynamics approaeh often leads to models of supply chains described in the form of a dynamical system. Dynamical systems theory provides a powerful framework for rigorous analysis of sueh models and thus can be used to supplement the system dynamics simulation approach. We illustrate this in section 5. using some nonlinear models, which consider the elTect of priority, heterogeneity, feedback, delays and resource sharing on the performance of supply chains. Furthermore, the large volumes of data generated from simulations ean be used to understand and comprehend the emergent dynamics of supply ehains. Even though an exaet understanding of the dynamics is difficult in complex systems, archetypal behaviour patterns can often be recognized, using techniques from complexity theory like Nonlinear Time Series Analysis and Computational Mechanics, which are discussed in section 6. The benefits of integrated supply chain concepts are widely recognized, but the analytical tools that can exploit those benefits are scarce. In order to study supply ehains as a whole, it is critical to understand the interplay of organizational structure and functioning of supply chains. Network dynamics, an extension of nonlinear dynamics to networks, provides a systematic framevt'ork to deal with sueh issues and is discussed in section 7. We conclude in section 8, with the reeommendations for future research.

2. Complex adaptive systems Many natural systems, and inereasingly many artilieial (man-made) systems as well. are characterized by apparently complex behaviours that arise as the result of nonlinear spatio-temporal interactions among a large number of components or subsystems. We use the term a^ent and luxle interchangeably to refer to the eomponent or subsystems. Examples of sueh natural systems include immune systems, nervous systems, multi-cellular organisms, ecologies, inseet societies and social organizations. However, such systems are not confmed to biology and society. Engineering theories of controls, communications and cotnputing have matured in recent decades, facilitating the creation of large-scale systems which have turned out to possess bewildering complexity, almost equivalent to that of biological systems. Systems sharing this property inelude parallel and distributed computing systems, communication networks, artificial neural networks, evolutionary algorithms, large-scale software systems, and economies. Such systems have been commonly referred to as Complex Sy.slems (Baranger 2005. Bar-Yam 1997. Adami 1998. Flake 1998). However, at the present time, the notion of a complex system is not precisely delineated. The most remarkable phenomenon exhibited by the complex systems is the emergence of highly structured collective behaviour over time from the interaction of simple subsystems without any centralized control. Their typical charaeteristies include: dynamics involving interrelated spatial and temporal effects, eorrelations over long length and timescales, strongly coupled degrees of freedom.

4238

A . Suraiui et al.

non-interchangeable system elements. They exist in quasi-equilibrium and show a combination of regularity and randomness (i.e. interplay of chaos and non-chaos). Such systems have structures spanning several scales and show emergent behaviour. Emergence is generally understood to be a process that leads to the appearanee of structure not directly described by the defining constraints and instantaneous forces that eontrol a system. The combination of structure and emergence leads to self-organization, which is what happens when an emerging behaviour has an efi!"eet of changing the structure or creating a new structure. CAS is a special category of complex systems to accommodate living beings. As the name suggests, they are capable of changing themselves to adapt to changing environment. In this regard, many artifieial systems like those slated earlier can be considered as CAS, due to their capability of evolving. Coexistence of competition and cooperation is another dichotomy exhibited by CAS. A CAS can be considered as a network of dynamieal elements where the states of both the nodes and the edges can change, and the topology of the network itself often evolves in time in a nonlinear and heterogeneous fashion. A dynamical system can be considered as simply "obeying the laws of physics". From another perspective, it ean be viewed as processing information: how systems obtain information, how they incorporate that information in the models of their surroundings, and how they make decisions on the basis of these models determine how they behave (Llyod and Slotine 1996). This leads to one of the more heuristic definitions of a complex system: one that 'stores, processes and transmits, information" (Sawhil 1995). From a thermodynamic viewpoint, sueh systems have the total energy (or its analogy) unknown, yet something is known about the internal state structure. In these large open systems (that do not possess well-defined boundaries), energy enters at low entropy and is dissipated. Open systems organize largely due to the reduction in the number of active degrees of freedom eaused by dissipation. Not all behaviours or spatial configurations can be supported. The result is a hmitation of the collective modes, cooperative behaviours, and coherent struetures that an open system can express. A central goal of the sciences of complex systems is to understand the laws and mechanisms hy whieh complicated, coherent global behaviour can emerge from the collective activities of relatively simple, locally interacting components. Complexity arises in natural systems thorough evolution, while design plays an analogous role for the complex engineering systems. Convergent evolution/ design leads to remarkable similarities at a higher level of organization, though at the molecular or device level, natural and man-made systems differ significantly. Complexity in both cases is driven far more by the need for robustness to uncertainty in the environment and eomponent parts than by basic functionality. Through design/evolution, such systems develop highly structured, elaborate internal configurations, with layers of feedback and signalling. Protocols organize highly structured and complex modular hierarchies to achieve robustness, but also create fragilities stemming from rare or ignored perturbations. The evolution of protocols ean lead to a robustness/eomplexity/fragility spiral where complexity added for robustness also adds new fragilities, which in turn leads to new and thus spiralling complexities (Csete and Doyle 2002). However, all this complexity remains largely hidden in normal operation and only becomes conspicuous when contributing to rare cascading failures or through chronic fragility/complexity evolutionary spirals. Highly Optimized Tolerance (HOT) (Carlson and Doyle 1999)

Supply-chain iwinorks: a complex Uihiplivc .sy.sWms pcispcclive

4239

has been introduced recently to rocus on the "robust, yet fragile' nature of complexity. It is also beeoming increasingly clear that robustness and complexity in biology. ecology, technology, and social systems are so intertwined that they must be treated in a unified way. Given the diversity of systems falling into this broad class, the discovery of any commonalities or 'universal' laws underlying sueh systems requires a very general theoretical framework. The scientific study of CAS has been attempting to find common characteristics and/or formal distinctions among complex systems that might lead to a better understanding of how complexity develops, whether it follows any general scientific laws of nature, and how it might be related to simplicity. The attractiveness of the methods developed in this research eflbrt for general-purpose modelling, design and analysis lies in their ability to produce complex emergent phenomena out of a small set of relatively simple rules, constraints and the relationships couched in either quantitative or qualitative terms. We believe that the tools and techniques developed in the study of CAS offer a rich potential for design, modelling and analysis of large-scale systems in general and supply chains in particular.

3. Supply-chain networks as complex adaptive systems A supply-chain network transfers information, products and finances between various suppliers, manufacturers, distributors, retailers and customers. A supply chain is characterized by a forward flow of goods and a backward flow of information. Typically, a supply chain comprises two main business processes: material management and physical distribution {Min and Zhou 2002). The material management supports the complete cycle of material flow from the purchase and internal control of production material to the planning and control of work-in-process, to the warehousing, shipping, and distribution of finished products. On the other hand, physical distribution encompasses all the outbound logisties activities related to providing customer services. Combining the activities of material management and physical distribution, a supply chain represents not only a linear chain of one-on-one business relationships but a web of multiple business networks and relationships. Supply-chain networks contain an emergent phenomenon. From the view of eaeh individual entity, the supply chain is self-organizing. Although the totality may be unknown, individual entities partake in the grand establishment of the network by engaging in their localized decision-making, i.e. in doing their best to select capable suppliers and ensure on-time delivery of products to their buyers. The network is characterized by nonlinear interactions and strong interdependencies between the entities. In most circumstances, order and control in the network are emergent, as opposed to predetermined. Control is generated through nonlinear though simple behavioural rules that operate based on local information. We argue that a supplychain network forms a complex adaptive system: Structures .spanning several .scakw. The supply-chain network is a bi-Ievel hierarchical and heterogeneous network where, at the higher level, each node represents an individual supplier, manufacturer, distributor, retailer or customer. However, at the lower level, the nodes represent the physical entities that exist inside each node in the upper level. The heterogeneity of

4240

A. Surana cl ill.

most networks is a function of various technologies being provided by whatever vendor could supply them at the time their need was recognized. Strongly coupled degrees of freedom and eorrelalions over long length and lime.srales: Different entities in a supply chain typieally operate autonomously with different objectives and subject to different set of constraints. However, when it comes to improving due date performance, increasing quality or redueing costs, they become highly inter-dependent. It is the flow of material, resources, information and finances that provides the binding force. The welfare o\' any entity in the system directly depends on the performanee of the others and their willingness and ability to coordinate. This leads to eorrelations between entities over long length and timeseales. Coexistence of competition and cooperation: The entities in a supply chain often have conflicting objectives. Competition abounds in the form of sharing and contention of resources. Global eontrol over nodes is an exception rather than a rule; more likely is a localized cooperation out of which a global order emerges, whieh is itself unpredictable. Nonlinear dynamics involving interrelated spatial and temporal effects: Supply chains have a wide geographic distribution. Customers can initiate transactions at any time with little or no regard for existing load, thus contributing to a dynamic and noisy network character. The charaeteristies of a network tend to drift as workloads and configuration change, producing a nonstationary behaviour. The coordination protocols attempt to arbitrate among entities with resource conflicts. Arbitration is not perfect, however; hence, over- and under-correetions contribute to the nonlinear character of the network. Quasi-equilibrium and combination of regularity and randomness (i.e. interplay of chaos and non-ehaos): The general tendency of a supply ehain is to maintain a stable and prevalent configuration in response to external disturbances. However, they can undergo a radical structural change when they are stretched from equilibrium. At such a point, a small event can trigger a cascade of changes that eventually can lead to system-wide reconfiguration. In some situations, unstable phenomena can arise, due to feedback structure, inherent adjustment delays and nonlinear decision-making processes that go in the nodes. One of the eauses of unstable phenomena is that the information feedback in the system is slow relative to the rate of changes that occur in the system. The first mode of unstable behaviour to arise in nonlinear systems is usually the simple one-cycle self-sustained oscillations. If the instability drives the system further into the nonlinear regime, more complicated temporal behaviour may be generated. The route to chaos through subsequent period-doubling bifurcations, as certain parameters of the system are varied, is generie to a large class of systems in physics, chemistry, biology, economies and other fields. Functioning in a ehaotic regime deprives the ability for long-term predictions about the behaviour of the system, while short-term predictions may be possible sometimes. As a result, control and stabilization of such a system beeome very difficult. Emergent helwviow and self-organization: With the individual entities obeying a deterministic selection proeess. the organization of the overall supply chain emerges through a natural process of order and spontaneity.

Siipplv-cliaiii networks: a complex adaptive sy.\teins perspective

4241

This emergence of highly structured collective behaviour over time from the interaclion of ihe simple entities leads to fulfilment of customer orders. Demand amplification and inventory swing arc two of the undesirable emergent phenomena that can also arise. For instance, the decisions and delays downstream in a supply chain often leads to amplifying a non-desirable effect upstream, a phenomenon commonly known as the "bull whip" effect. .Adaptation and evo/titiotr. A supply chain reacts to the environment and thereby creates its environment. Operationally, the environment depends on the chosen scale of analysis, e.g. it can be taken as the customer market. Typically, significant dynamism exists in the environment, which necessitates a constant adaptation of the supply network. However, the environment is highly rugged, making the eo-evolution difficult. The individual entities constantly observe what emerges from a supply network and adjust their organizational goals and supporting infrastructure. Another common adaptation is through altering boundaries of the network. The boundaries ean change as a result of including or excluding particular entity and by adding or eliminating connections among entities, thereby changing the underlying pattern of interaction. As we discuss next, supply-chain management plays a critical role in making the network evolve in a coherent manner.

3.1 Supply-chain management Supply-chain management is the integration of key business processes from endusers through original suppliers that provide produets. services, and information and add value for customers and other stakeholders (Cooper et ai 1997). It involves balancing reliable customer delivery with manufacturing and inventory costs. It is evolved around a customer-focused corporate vision, which drives changes throughout a firm's internal and external linkages and then captures the synergy of interfunctional, inter-organi/ational integration and coordination. Owing to the inherent complexity, it is a challenge to coordinate the actions of entities across organizational boundaries so that they perform in a coherent manner. An important element in managing supply-chain networks is to control the ripple effect of lead time so that the variability in supply chain can be minimized.
Suppliers Manufacturers Distributors Retailers Customers

Figure I.

Supply-chain network.

4242

A. Surana ct al.

Demand forecasting is used to estimate demand for each stage, and the inventory between stages for the network is used for protecting against fluctuations in supply and demand across the network. Owing to the decentralized control properties of the SCN. control of ripple elTect requires coordination between entities in performing their tasks. With the increase in the number of participants in the supply chain, the problem of coordination has reached another dimension. Two important organizational and market trends that are on their way have been the atomization of markets as well as that of organizational entities (Balakrishnan ('/ al. 1999). In such a scenario, the product realization process has a continuous customer involvement in all phases from design to delivery. Customization is not only limited to selecting from pre-determined model variants; rather, product design, process plans, and even the supply chain configuration have to be tailored for each customer. The product-realization organization has to form on the Hy, as a consortium of widely dispersed organizations to eater to the needs of a single customer. Thus, organizations consist of series of opportunistic alliances among several focused organizational entities to address particular market opportunities. For manufaeturing organizations to operate effectively in this environment of dynamic. virtual alliances, products must have modular architectures, processes must be well characterized and standardized, documentation must be widely accessible, and systems must be interoperable. Automation and intelligent information processing is vital for diagnosing problems during product realization and usage, coordination, design and production schedules, searching for relevant information in muhi-media databases. These trends exaeerbate the challenges of coordination and collaboration as the number of product realization networks increase, and so does the number of partners in eaeh network. Building a larger inventory can be used as a general means for dealing with highly changing market demand and short-life-cycle products. However, augmenting inventory building with information may be a useful approach. Information about the material lead time from different suppliers can be used for planning the material arrival, instead of simply building up an inventory. The demand information can be transmitted to the manufacturers on a timely basis, so that the orders can be fulfilled with less inventory costs. In fact, it is widely realized that the successful integration of the entire supply-chain process depends heavily on the availability of accurate and timely information that can be shared by all members of the supply chain. Supply-chain management now increasingly relies on information technology, as discussed below.

3.2 Information technohfiy in supply-chain management Information technology, with its capability of providing global reaeh and wide range of connectivity, enterprise integration, micro-autonomy and intelligence. object and networked-oriented computing paradigms and rich media support, has been the key enabler for the management of modern manufacturing enterprises (Balakrishnan et al. 1999). It is vital for eliminating collaboration and coordination costs, and permits the rapid setup of dynamic information exchange networks. Connectivity permits involvement of customers and other stakeholders in all aspects of manufaeturing. Enterprise integration facilitates seamless interaction

Supply-chain nciworks: a complex aiiaplivv svMems pcr.spccfive

4243

among global partners. Micro-autonomy and intelligence permit atomic tracking and remote control. New software paradigms enable distributed, intelligent and autonomous operations. Distributed computing facilitates quick localized decisions without losing the vasl data-gathering potential und powerful computing capabilities. Rich media support, which includes capabilities like digitization, visualization tools and virtual reality, facilitate collaboration and immersion. Many improvements have occurred in supply-chain management because IT enables dynamic changes in inventory management and production, and it assists the managers in coping with uncertainty and lead time through improved collection and sharing of information between supply-chain nodes. The success of an enterprise is now largely dependent on how Its information resources are designed, operated and managed, especially with the information technology emerging as a critical input to be leveraged for significant organizational productivity. However, it is difficult to design an information system that can handle the information needs of supply-chain nodes to allow cRicienl. flexible and decentralized supply-chain management. The main hurdle in efficiently using information technology is the lack of our understanding of the organizational, functional and evolutionary principles of supply chains. Recognizing supply chains as CAS can. however, lead lo novel and effective ways to understand their emergent dynamics. It has been found that many of the diverse looking CAS share similar chiu'acteristics and problems, and thus can be tackled through similar approaches. While, at present, networks are largely controlled by humans, the complexity, diversity and geographic distribution of the networks make it necessary for networks to maintain themselves in a sort of evolutionary sense, just as biological organisms do {Maxion 1990). Similarly, the problem of coordination, which is a challenge in supply chains, has been routinely solved by biological systems for literally billions of years. We believe that the complexity, flexibility and adaptability in the collective behaviour oi^ the supply chains can be accomplished only by importing the mechanisms that govern these features in nature. Along with these robust design principles, we require equally sound techniques for modelling and analysis of supply chains. This forms the focus of this paper. We first give a brief overview of the main techniques that have been used for modelling and analysis of supply chains, and then discuss how the science of complexity provides a genuine extension and reformulation of these approaches.

4. Modeiling and analysis of suppiy-chain networks As pointed out. the key challenge in designing supply-chain networks or. for that matter, any large-scale systems is the difficulty of reverse engineering, i.e. determining what individual agent strategies lead to the desired collective behaviour. Because of this difficulty hi understanding the effecl of individual characteristics on the coiloctive behaviour of the system, simiilalion has been the primary tool for designing and optimizing such systems. Simulation makes investigations possible and useful when, in the real-world situation, experimentation would be too costly or. for ethical reasons, not feasible, or where the decisions and their consequences are well separated in space and lime, it seems at present that large-scale siinulalions

4244

A. Swana et al.

of future complex processes may be the most logical, and perhaps an important vehicle to study them objectively (Ghosh 2002). Simulation in general helps one to detect design errors, prior lo developing a prototype in a cost-effective manner. Second, simulation of system operations may identify potential problems that might occur during actual operation. Third, extensive simulation may potentially detect problems thai are rare and otherwise elusive. Fourth, hypothetical concepts lliat do nol exisl in nature, even those that defy natural laws, may be studied. The increased speed and precision of today's computers promise the devclopmcnl of high-fidelity models of physical and natural processes, models that yield reasonably accurate results, quickly. This In turn would permit system architects to study the performance impact of a wide variation of key parameters quickly and. in some cases, even in real time. Thus, a qualitative improvement in system design may be achieved. In many cases, unexpected variations in external stress can be simulated quickly to yield appropriate system parameters values, which are then adopted Inlo the system to enable it to successfully counteract the external stress. Mathematical analysis, on the other hand, has to a play a critical role because it alone can enable us to formulate rigorous generalizalions or principle. Neither physical experiments nor computer-based experiments on their own can support such generalizations. Physical experiments usually are limited to supplying inputs and constraints for rigorous models, because experiments themselves are rarely described in a language that permits deductive exploration. Computer-based experiments or simulations have rigorous descriptions, but they deal only in specifics. A well-designed mathematical model, on the other hand, generalizes the particulars revealed by the physical experiments, computer-based models and any interdisciplinary comparisons. Using mathematical analysis, we can study the dynamics, predict long-term behaviour, and gain insights into system design: e.g. what parameters determine group behaviour, how individual agent characteristics affect the system and that the proposed agent strategy leads to the desired group behaviour. In addition, mathematical analysis may be used to select parameters that optimize a system's collective behaviour, prevent instabilities, etc. It seems that successful modelling efforts of large-scale systems like supply-chain networks, large-scale software systems, communication networks, biological ecosystems, food webs, social organizations, etc. would require a solid empirical base. Pure abstract mathematical contemplation would be unlikely to lead to useful models. The discipline of physics provides an appropriate parallel; advances in theoretical physics are more often than not inspired by experimental findings. The study of supply-chain networks should therefore involve an amalgam of both simulailon and aiuilyiical techniques. Considering the broad spectrum of a supply-chain, no model ean capture all the aspects of supply-chain processes. The modelling proceeds at ihree levels: 1. competitive stralegic analysis, which includes location-allocation decisions, demand planning, distribution channel planning, strategic alliances, new product development, outsourcing. IT selection, pricing and network structuring; 2. taclical problems like inventory control, production/distribution coordination, material handling and layout design;

Supply-chain network.s: a complex uduplivc sy.stcm.'i pcr.spcclive

4245

3. operational level problems, which include routing/scheduling, workforce scheduling and packaging. The individual models in supply chains can be categorized into four classes (Min and Zhou 2002): 1. deterministic: single objective and multiple objective models; 2. stochastic: oplimal conlrol theoretic and dynamic programming models; 3. hybrid: with elements of both delerminislic and stochastic models and includes inventory theoretic and simulations models; 4. IT-driven: models that aim to integrate and coordinate various phases of supply-chain planning on a real-time bases using application software. like ERP. Mathematical programming techniques and simulation have been two approaehes for the analysis and study of the supply-ehain models. Mathematical programming mainly takes into eonsideration static aspeets of the supply chain. Simulation, on the other hand, studies dynamics in supply chains and generally proceeds based on "system dynamics' and "agent-based" methodologies. System dynamics is a continuous simulation methodology that uses concepts from engineerhig feedback control to model and analyse dynamic socio-economic systems (F"orrester 1961). The mathematical description is realized with ordinary difierential equations. An important advantage of system dynamics is the possibility to deduce the occurrence of a specific behaviour mode because the structure that leads to the system dynamics is made transparent. We present some nonlinear models in section 5 which are useful for understanding the complex interdependencies, effects of priority, nonlinearities. delays, uncerlainlies and competilion/cooperation for resource-sharing in supply chains. The drawback of system dynamics models is ihat the structure has to be determined before starting the simulation. Agentbased modelling (a technique from complexity theory), on the other hand, is a "bottom up approach" which simulates the underlying processes believed responsible for the global pattern, and allows us to evaluate what mechanisms are most influential in producing thai emergent pattern. In Schieritz and Grobler (2003), a hybrid modelling approach has been presented thai intends to make the system dynamics approach more flexible by combining il with the discrete agent-based modelling approach. Such large-scale simulations with Iheir many degrees of freedom raise serious technical problems about the design of experiments and the sequence in which they should be carried out in order to obtain the maximum relevant information. Furthermore, in order to analyse data from such large-scale simulations, we require systematic analytieal and statistical methods. In section 6, we describe two such techniques: nonlinear lime scries analyses and a>mpulali(inal mechanics. A useful paradigm for modelling a supply chain, taking into consideration the detailed pallcrn of interaction, is to view it as a network. A network is essentially anything that can be represented by a graph: a set of points (also generJcally called nodes or vertices), connected by links (edges, ties) representing some relationship. Networks are inherently difficult to understand due to their structural complexity, evolving structure, connection diversity, dynamical complexity of nodes, node diversity and nieta-complication where all these factors influence each other.

4246

A. Surana el al.

Queuing theory has primarily been used to address the steady-state operation of a lypical network. On the other hand, techniques from mathematical programming have been used to solve the problem of resource allocation in networks. This is meaningful when dynamic transients can be disregarded. However, present-day supply-chain networks are highly dynamic, reconfigurable. intrinsically nonlinear and non-stationary. New tools and techniques are required for their analysis sueh that the structure, function and growth of networks can be considered simultaneously. In this regard, we discuss "network dynamics' in section 7. which deals with such issues and can be used to study the structure of supply chain and its implication for its functionality. Understanding the behaviour of large complex networks is the next logical step for the iicid of nonlinear dynamics, because they are so pervasive in the real world. We begin with a brief introduction to dynamical systems theory, in particular nonlinear dynamics in next section.

5. Dynamieal systems theory Many physical systems that produce conlinuous-time response can be modelled by a set of differential equations of the form: ^=./(.v.). (1)

where y = (_V|(O..i':(O .''-((O) represents the state of the system and may be thought of as a point in a suitably defined space S, which is known as pha.se spuce, and t/ (W|(/),72(/), a,,,{t)) is a parameter vector. The dimensionality of 5 is the number of a priori degrees of freedom in the system. The vector field /'{ r, c/) is in general a nonlinear operator acting on points in S. If./(.v. (/) is locally Lipschtiz, the above equation defines an initial value problem in the sense that a unique solution curve passes through each point i- in the phase space. Formally, we may write the solution at time / given an initial value V as y(t) = cp,y{). (p, represents a oneo parameter family of maps of the phase space into itself. We can perceive the solutions to all possible initial value problems for the system by writing them collectively as (PiS. This may be thought of as a flow of points in the phase spaee. Initially, the dimension of the set <p,S will be that of S itself. As the system evolves, however, it is generally the case for the dissipativc system that Ihe flow contracts on to a set of lower dimension known as an attractor. The attractors can vary from simple stationary, limit eyele. quasi-periodic to complicated chaotic attractors (Strogatz 1994, Ott 1996). The nature of attractor changes as parameters (a) is varied, a phenomenon studied in bifurcation analysis. Typically, a nonlinear system is always chaotic for some range of parameters. Chaotic attractors have a structure that is not simple; they are often not smooth manifolds and frequently have a highly fractured structure, which is popularly referred to as Fractals (self-similar geometrical objects having a slruclure at every scale). On this attractor, stretching and folding characterize the dynamics. The stretching phenomenon causes the divergence of nearby trajectories and the folding phenomenon constraints the dynamics to finite region of the state space. This accounts for the fractal structure of attractors and the extreme sensitivity to changes in initial conditions, which is a hallmark of chaotic behaviour. A system under chaos is unstable everywhere and never settles down, producing

Siipplv-cliiiin ncUvork.K: a complex adaptive systems perspective

irregular and aperiodic behaviour, which leads to a continuous broadband spectrum. While this feature can be used to distinguish chaotic behaviour from siciHonary. Until cycle, quasi-perioclic motions using standard Fourier analysis, it makes it difficult to separate it from noise which also has a broadband spectrum. It is this 'deterministic randomness' of chaotic behaviour which makes standard linear modelling and prediction techniques unsuitable for analysis.

5.1 Noniinear models for the supply chain Understanding the complex interdependencies. effects of priority, nonlinearities. delays, uncertainties and competition; cooperation for resource sharing are fundamental for prediction and control of supply chains. A system dynamics approach often leads to models of supply chains, which can be described in the form of equation (1). Dynamical systems theory provides a powerful framework for rigorous analysis of such models and thus can be used to supplement the system dynamics approach. We next describe some nonlinear models and their detailed analysis. These models can be used either to represent entities in a supply chain or as macroscopic models, which capture collective behaviour. The models reiterate the fact that simple rules can lead to complex behaviour, which in general are difficult to predict and control.

5.1.1 Pre-emptive queuing model with delays. Priorily and heterogeneity are fundamental to any logistic planning and scheduling. Tasks have to be prioritized in order to do the most important things first. This comes naturally as we try to optimize an objective and assign the tasks their 'importance'. Priorities may al.so arise due to the non-homogeneity of the system where the "knowledge" level of one agent is different from the other. In addition, in all logistics systems, resources are limited, in both

B,

AFigure 2. Pre-emptive queuing model.

4248

A. Surana et al.

time and spaee. Temporal dependence plays an important role in logistic planning (inlerdependeney). Sometimes, they can also arise from the physical facts when different stages of processing have certain temporal constraints. The considerations regarding the generality of assumptions and the clear oneto-one correspondence between the physical logistics tasks and the model parameters described in (Erramilli and Forys 1991) made us apply their queuing model in the context of supply chains (Kumara el al. 2003). The queuing system eonsidered here has two queues (A and B) and a single server with the following characteristics: once served, the class A customer returns as a class B customer after a constant interval of time; Class B has non-pre-emptive priority over class A, i.e. the class A queue is not served until the class B queue is emptied: the schedules are organized every Tunits of time, i.e. if the low priorily queue is emptied within time T. the server remains idle for the reminder of the interval; finally, the higher-priority class B has a lower service rate than the lowpriority class A. Suppose the system is sampled at the end of every schedule cycle, and the following quantities are observed at the beginning of the klh interval: A/,: queue length of low-priority queue; 5^: queue length of high-priority queue; C/,\ outflow from low-priority queue in the Alh interval; >A: outflow from high-priority queue in the A'th interval: ?./,'. inflow to low-priority queue from the outside in the klh interval. The system is characterized by the following parameters: /i,; rate per unit of the schedule cycle at which the low-priority queue can be served; it/,: rate per unit of the schedule cycle at which the high-priority queue can be served; /: the feedback interval in units of ihe schedule cycle. The following four equations then completely describe the evolution of the system:
/(,^, = / 1 A + A , - C , (2)

(3)

A, ^m\n{Bi,+C\_i,iii,).

(5)

Equations (2) and (4) are merely conservation rules, while equations (3) and (5) model the constraints on the outflows and the interaction between the queues. This model, while conceptually simple, exhibits surprisingly complex behaviours. The analytic approach to solve for the flow model under constant arrivals (i.e. A^ = X for all k) shows several classes of solutions. The system batches its workload even lor perfectly smooth arrival patterns. The characteristics of the behaviour of the system are as follows: 1. Above a threshold arrival rate (A. > /i/,/2). a momentary overload can send the system into a number of stable modes of oscillations. 2. Each mode of oscillations is characterized by distinct average queuing delays.

Supply-chain networks: a complex adaptive systems perspective

4249

3. The extreme sensitivity to parameters, and the e.\istence of chaos, implies thai the syslem at a given time may be any one ot' a number o^ distinct steady-state modes. The batching of the workload can cause significant queuing delays, even at moderate occupancies. Also, such oscillatory behaviour significantly lowers the real-time capacity of the system. For details of the application of this model in a supply-ehain context, refer to Kumara et al. (2003).

5.1.2 Managerial systems. Decision-making is another typical characteristic in which the entities in a supply chain arc continuously engaged. Entities make decisions to optimize their self-interests, often based on local, delayed and imperfect information. To illustrate the elfects of decisions on the dynamics of supply chain as a whole, we consider a managerial system which allocates resources to its production and marketing departments in accordance with shifts in inventory and or backlog (Rasmussen and Moseklide lySH). It has four level variables: resources in production, resources in sales, inyentory of finished products and nutnher <if customers. In order to represent the time required to adjust production, a third-order delay is introduced between production rate and inventory. The sum of the two resource variables is kept constant. The rate of production is determined from resources in production through a nonlinear function, which expresses a decreasing productivity of additional resources as [he company approaches maximum capacity. The sales rate, on the other hand, is determined by the number of customers and by the average sales per customer-year. Customers are mainly recruited through visits of the company salesman. The rate of recruitment depends upon the resources allocated to marketing and sales, and again it is assumed that there is a diminishing return to inereasing sales activity: once recruited, customers are assumed to remain with the company for an average period AT. the as.sociation time. A difference between production and sales causes the inventory to change. The Company is assumed to respond to such changes by adjusting its resource allocation. When the inventory is lower than desired, on the other hand, resources are redirected from sales to production. A certain minimum of resources is always maintained in both production and sales. In the model, this is secured by means oi two limiting factors, which reduce the transfer rate when a resource floor is approached. Finally, the model assumes that there is a feedback from inventory to customer defection rate. If the inventory of finished products becomes very low. the delivery time is assumed to become unacceptable to many customers. As a consequence, the defection rate is enhanced by a factor 1 + //. The managerial system described is controlled by two interacting negative feedback loops. Combined with the delays involved in adjusting production and sales, these loops create the potential for oscillatory behaviour. If the transfer of resourees is fast enough, this behaviour is destabilized, and the system starts to perform selfsustained oscillations. The amplitude of these oscillations is finally limited by the various nonlinear restrictions in the model, particularly by the reduction in resource transfer rate as lower limits to resources in production or resources in sales are approached.

4250

.4. Suniiw cl al.

Customer Defection Riite

AT

Sales Per Ciistoiner Year

Figure 3. Managerial system.

A series of abrupt changes in the system behaviour is observed as competition between the basic growth tendency and nonlinear limiting factors is shifted. The simple one-cycle attractor corresponding to // 10, becomes unstable for //13, and a new stable attractor with twice the original period arises. If H is increased to 28. the stable attractor attains a period of 4. As H is further increased, the period-doubling bifurcations continue until /y=30. the threshold to chaos, is exceeded. The system now starts to behave in an aperiodic and apparently random behaviour. Hence, the system shows chaotic behaviour through a series of perioddoubling bifurcations. 5.1.3 Deterministic qut'uinj^ model. In this section, we consider an alternate discrete-lime deterministic queuing model, for studying decision-making at an entity level in supply chains. The model consists of one server and two queuing

Supply-chain nciwoik.s: a complex adaptive sy.'ilems perspective

4251

lines {X and Y) representing some activity {Feichtinger ct al. 1994). The input rates of both queues are constant, and their sum equals the server capacity. In each time period, the server has to decide how much time to spend on each of the two activities. The following quantities can be defined: a: constant inpul rate for activity X\ fi: constant inpul rale for activity Y: ^y. timt; spenl on activity A"; Oy-: time spent on activity Y: .Y^: queue length of .V; y^: queue length of Y. The amount of time <t>x and <t)y that will be spent on activities X and Y in period A--1-1 are determined by an adaptive feedback rule depending on the difference of the queue lengths \\ and i^. The decision rule or policy function slates that longer queues are served with a higher priority. Two possibilities considered arc: 1. All-or nothing ilecision: The server decides to spend all its time on the activity corresponding to the longer queue. Hence, O is a Heaviside function given by
0 ( . v - \) = 1
= 0 if .V > V

if.v < V.

(6)

2. Mixed sohitiotts: The server decides to spend most of its time to the activity corresponding lo the longer queue. For this decision function, an S-shaped logistic function is used as given by:

The parameter k tunes Ihe "steepness' of the S-shape. With these decision functions, the new queue lengths .V/,-+| and equations

are given

(8)

Figure 4. Delerminislic queuing model.

4252

A. Surana ct a!.

Using the constraints a-^fi=l and <J'A-i-'I'j 1- it is sufficient to consider the dynamics of the map in order to study the behaviour of the system -2). (9)

For 0 < A < 4 and for all 0 < Q ' < I, ihc map / has a globally stable equilibrium. Simulation shows that when the parameter k is not too large, the bifurcation diagrams with respect to a are simple. For larger values of k (e.g. A' 7.3). chaotic behaviour arises after infinitely many period doubling bifurcations, as a is increased from 0.0 to 0.3. However, when cr is further increased from 0.3 to 0.5, chaos disappears after many period-halving bifurcations. For 0.5 < a < 1. the bifurcations scenario is qualitatively the same as for 0<ff<0.5. since the system is symmetric w.r.t. a = 0.5 and .\ = 1. Physically, when a is close to 0. there is stable equilibrium, meaning that in the long run., in each time period, the server spends a fixed proportion of time to each of the two activities, and it spends most of the time on the activity K with the highest input rate. For a close to 1, we have same behaviour, with the activities X and Y interchanged. For a close to 0.5. i.e. when the input rates of the two activities are almost equal, the equilibrium is unstable, and there is stable period 2 orbit. This means that in one period, most of the time is spent on activity X, and then in the next period most of the time is spent on activity Y. and again on activity X and so on. Chaotic behaviour arises when a is somewhere between 0 and 0.5 or between 0.5 and I, for a 1/3 and a 2/?i. Hence, a steep decision function together with a situation where the input rate of one activity is around twice the input rate of the other activity leads to irregular queue lengths. As k -^ DC. the decision function <t> converges to the Heaviside function. The dynamical behaviour of the queuing model in that case is equivalent to rigid rotation on a circle. For rational a = p:q of the input rate, every point .v is periodic with period (/. In that case, for all q time periods, p time periods are completely spent on the first activity, while the remaining q p time periods are spent on the other activity. On the other hand, when a is irrational, the dynamical behaviour is quasi-periodic, and every point .v is aperiodic.

5.2 Dynamical models of resource allocation: Computational ecosystems Because of limited resources, resource sharing and allocation is a fundamental problem in any supply chain. The manner in which the resources are shared and utilized has a significant impact on the performance of a supply chain. It also dictates how cooperation/competition arises and is sustained in a supply chain. Resources can be of various types: physical resources, manpower, information and monetary. With the IT architectures being developed to realize supply chains, sharing of computational resources (like CPU, memory, bandwidth, databases, etc.) is also becoming a critical issue. It is through resource sharing that interdependencies arise between different entities. This leads to a complex web ot' interactions in supply chains just like in a food web or ecology. As a result, such systems can be referred to as 'Computational Ecosystems" (Hogg and Huberman 1988) in analogy with biological ecosystems. 'Computational Ecosystems" is a generic model of the dynamics of resource allocation among agents trying to solve a problem collectively. The model captures the following features: distributed control, asynchrony in execution, resource contention and cooperation among agents and concomitant problem of incomplete

Supply-chain nclwork.s: a complex at/tiptivc systems perspective

4253

knowledge and delayed informalion. The behaviour of each agent is modelled using a payoff function whose nature determines whether an agent is cooperative or competitive. The agent here can be any entity in a supply chain like a distributor, retailer, etc. or a software agent in an e-commerce scenario. The state of the system is represented as an average number of entities using different resources and follows a delay differential equation under a mean Held approximation. The resources can he physical or computational as discussed before. For example, in case of two resources with n identical agents, the rate of change of occupation of a resource is given by: (10) where. (i(/)) is the expected number of agents using resource 1 at a given instant of time /; a is the expected number of choices made by an agent per unit time; p is a random variable that denotes that resource 1 will be perceived to have a higher payotr than resource 2; and (p) gives its expected value. The global performance of the ecosystem can be obtained from the above equation. Under ditTerent conditions of delay, uncertainty, and cooperation/competition, the system shows a rich panoply of behaviours ranging from stable, sustained oscillations to intermittent chaos and finally to fully developed chaos. Furthermore. the following generic deductions can be made from this model (Kephart ct al. 1989): while information delay has an adverse impact on the system performance, uncertainty has a profound effect on the stability of the system. One can deliberately increase uncertainty in agents" evaluation of the merits of choices to make it stable but at the expense of performance degradation. A second possibility is a very slow re-evaluation rate of the agents, which makes them non-adaptive. Heterogeneity in

Model of Computational Ecosystems

Payoff E valuator Random Component

Payoff E valuator deterministic Component

Figtire 5. C'omptiiational ecosystems (r: lime delay; cr: standard deviation oC p).

4254

A. Surana et id.

the nature of agents can lead to more stability in the system compared with a homogenous case, but the system loses its ability to cope with une.xpected changes in the system such as new task requirements. On the other hand, a poor performance can be traced to the fact that the non-predictive agents do not take into account the information delay. If the agents are able to make accurate predictions of its current state, the information delay could be overcome, and the system would perform well. This results in a "co-evolutionary" system in which all oi' the individuals are simultaneously trying to adapt to one another. In such a situation, agents can act like Technical Analysts and System Analysts (Kephart et ai 1990). Agents as technical analysts (like those in market behaviour) use either linear extrapolation or cyclic trend analysis to estimate the current state of the system. On the other hand, agents as system analysts have knowledge about both the individual characteristics of the other agents in the system and how those characteristics are related to the overall system dynamics. Technical analysts are responsive to the behaviour of the system but suffer from an inability to take into account the strategies of other agents. Moreover, a good predictive strategy for a single agent may be disastrous if applied on a global scale. System analysts perform extremely well when they have very accurate information about other agents in the system but can perform very poorly when their information is even slightly Inaccurate. They take into accotmt the strategies of other agents but pay no heed to the actual behaviour of the system. This suggests combining the strengths of both methods to form a hybrid-adaptive system analyst, which modifies its assumptions about other agents in response to feedback about success of its own predictions. The resultant hybrid is able to perform well. In order to avoid chaos while maintaining a high performance and adaptability to unforeseen changes, more sophisticated techniques are required. One such way is by a reward mechanism (Hogg and Huberman 1991), whereby the relative number of computational agents following effective strategies is increased at the expense of the others. This procedure, which generates a right population diversity out of essentially homogenous ones, is able to control chaos by a series of bifurcations into a stable fixed point. In the above description, each agent ehooses among different resources according to its perceived payoff, which depends on the number of agents already using it. Even the agent with predictive ability is myopic, as it considers only its current estimate of the system state, without regard to the future. Expectations come into play if agents use past and present global behaviour in estimating the expected future payoff for each resource. A dynamical model of collective action that includes expectations ean be found in Glance (1993).

6. Models from observed data One of the central problems in a supply chain, closely related to modelling, is that of demand forecasting: given the past, how can we predict the future demand? The classic approach to forecasting is to build an explanatory model from first principles and measure the initial conditions. Unfortunately, this has not been possible for two reasons in systems like supply chains. Eirst. we still lack the general

Supply-chain networks: a complex culaprive systems perspective

4255

'first principles" for demand variation in supply chains, which are necessary to make good models. Second, due to the distributed nature of the supply chains, the initial data or the eonditions are often difficult to obtain. Because of these faetors. the modern theory of forecasting that has been used in supply chains views a time series A(/) as a realization of a random process. This is appropriate when effective randomness arises from complicated motion involving many independent, irreducible degrees of freedom. An alternative cause of randomness is chaos, which can occur even in very simple deterministic systems, as we discussed in the earlier sections. While chaos places a fundamental limit on longterm prediction, it suggests possibilities for short-term prediction. Random-looking data may contain only few irreducible degrees of freedom. Time traces of the state variable of such chaotic systems display a behaviour, which is intermediate between regular periodic or qtiasiperiodic motions, and unpredictable, truly stochastic behaviour. It has long been seen as a form of "noise" because the tools for its analysis were couched in language tuned to a linear process. The main such tool is Fotirier analysis, which is precisely designed to extract the composition of sines and cosines found in an observation .v(/). Similarly, the standard linear modelling and predictit>n techniques, such as autoregressive moving average (ARMA) models, are not suitable for nonlinear systems. With the advances in IT and science of complexity, both the challenges for forecasting can be revisited. Large-scale simulation and micro-autonomy (section 2) enable tracking of the detailed interaction between different entities in a supply chain. The large volumes of data thus generated can be used to understand demand patterns in particular and comprehend the emergence of other characteristics in general. Even though an exact prediction of future behaviour is difficult. often archetypal behaviour patterns can be recogni/ed using these data. Techniques from Ihe complexity theory like Nonlinear Time Series Analysis and Computational Mechanics are appropriate for this purpose, 6.1 Nonlinear lime-series analysis The need to extract interesting physical information about the dynamics of observed systems when they are operating in a chaotic regime has led to the development of nonlinear time .series analysis techniques. Systematically, the study of potentially, chaotic systems may be divided into three areas: identification of chaotie hchaviour, modelling and prediction and eontrol. The first area shows how chaotic systems may be separated from stochastic systems and, at the same time, provides estimates of the degrees of freedom and the complexity of the underlying chaotic system. Based on such results, identification of a state-space representation allowing for subsequent predictions may be carried out. The last stage, if desirable, involves control of a chaotic system. Given the observed behaviour of a dynamical system as a one-dimensional time series .Y(/J). we want to build models for prediction. The most important task in this process is phase space reconstruction, which involves building topologically and geometrically equivalent attractor. In general, steps in nonlinear time series analysis can be summarized as (Abarbanel 1996): Siffnal separation (/indini^ the sii^nal r. Separation of a broadband signal from broadband 'noise" using deterministic nature of signal.

4256

A. Surana et al. Pha.se space recon.slruction (finding ihe space): Using the method of delays, one can construct a series of vectors which is diffeomorphically equivalent to the attractor of the original dynamical system and at the same time distinguish il from the being stochastic. The basis for this is Taken's Embedding theorem (Takens 1981). Time-lagged variables arc used to construct vectors for a phase space in d/r dimension: V('O - [x{n)..\in +T) x(n + 1^//. - \)T)\. (II)

The time lag T can be determined using mutual information {Fraser and Swinney 1983) and UE using a false nearest-neighbours test (Kennel ct al. 1992). Cla.ssificaiion of ihc signal: System identification in nonlinear chaotic systems means establishing a set of invariants for each system of interest and then comparing observations wilh thai library of invariants. The invariants are properties of atlractor and are independent of any particular trajectory of the attractor. Invariants can be divided into two classes: fractal dimensions (Farmer el al. 1983) and Lyapunov exponents (Sano and Sawada 19H5). Fractal dimensions characterize ihe geometrical complexity of dynamics, i.e. how the sample of points along a system orbit are distributed spatially. Lyapunov exponents, on the other hand, describe the dynamical complexity, i.e. 'stretching and folding" in the dynamical process. Making models and prediction: This step involves determination of the parameters i/, of the assumed model of the dynamics:

yln+

1) ^ /( v(/0-'^'i-''2^

"/<*

which is consistent with invariant elassifiers (Lyapunov exponents, dimensions). The functional form F(-) often used includes polynomials, radial basis functions., etc. The Local False Nearest Neighbor (AbarbancI and Kennel 1993) test is used to determine how many dimensions are locally required to describe the dynamics generating the time series, without knowing the equations of motion, and hence gives the dimension for the assumed model. The methods for building nonlinear models are classified as Global and Local (Farmer and Sidorowich 1987, Casdalgi 1989). By dcfinilion. Local methods vary from point to point in the phase space, while Global Models are constructed once and for all in the whole phase space. Models based on Machine Learning techniques such as radial basis funetions or Neural Networks (Powell 1987) and Support Vector Machines (Mukherjee et al. 1997) carry features of both. They are usually used as global functional forms, but they clearly demonstrate localized behaviour, too. The techniques from nonlinear time series analysis are well suited for modelling the nonlinearities in the supply chains. For an application of nonlinear time series analysis in supply chains, the reader is referred to Lee et al. (2002). Using this, one can deduce that the time series is deterministic, so it should be possible in principle to build predictive models. The invariants can be used to effectively characterize the complex behaviour. For example, the largest Lyapunov exponent gives an indication of how far into ihe future reliable predictions can be made, while the fractal dimensions give an indication of how complex a

Siippiy-chain neinorks: a etnnptex adaptive systems perspective

4257

model should be chosen to represent the data. These models then provide the basis for systematically developing the control strategies. It should be noted that the funetional forms used for modelling in the fourth step above are eontinuous in their argument. This approach builds models viewing a dynamical system as obeying laws of physics. From another perspective, a dynamical system can be eonsidered as processing information. So. an alternative class of discrete 'computational' models inspired from the theory of automata and formal languages ean also be used for modelling the dynamics (Lind and Marcus 1996). 'Computational Mechanics" considers this viewpoint and describes the system behaviour in terms of its intrinsic computational architecture, i.e. how it stores and processes information. 6.2 Computational mechaiues Computational mechanics is a method for inferring the causal structure of stochastic processes from empirical data or arbitrary probabilistic representations. It combines ideas and techniques from nonlinear dynamics. Information theory and automata theory, and is, as it were, an 'inverse" to statistical mechanics. Instead of starting vi^ith a microscopic description of particles and their interactions, and deriving macroscopic phenomena, it starts with observed macroscopic data and infers the simplest causal structure: the 'f-machine" capable of generating the observations. The f:-machine in turn describes the system's intrinsic computation, i.e. how it stores and processes information. This is developed using the statistical mechanics of orbit ensembles, rather than focusing on the computational complexity of individual orbits. By not requiring a Hamiltonian, computational mechanics can he applied in a wide range of contexts, including those where an energy function for the system may not manifest as for the supply chains. Notions of complexity, emergence and self-organization have also been formalized and quantified in terms of various information measures (Shalizi 2005). Given a time series, the (unknowable) exact states of an observed system are translated into a sequence of symbols via a measurement channel (Crutchfield 1992). Two histories (i.e. two series of past data) carry equivalent information if they lead to the same (conditional) probability distribution in the future (i.e. if it makes no difference whether one or the other data series is observed). Under these circumstances, i.e. the effects of the two series being indistinguishable, they can be lumped together. This procedure identifies causal states and also identifies the structure of connections or succession in causal states and creates what is known as an 'epsilon-machinc'. The ^-machines form a special class of Deterministic Finite State Automata (DFSA) with transitions labelled with conditional probabilities and hence can also be viewed as Markov chains. However, finite-memory machines like f-machines may fail to admit a finite size model, implying that the number of casual states could turn out to be inlinite. In this case, a more powerful mode! than DFSA needs to be used. One proceeds by trying to use the next most powerful model in the hierarchy of machines known as the casual hierarchy (Crutchfield 1994), in analogy with the Chomsky hierarchy of formal languages. While >-machine reconstiuction" refers to the process of constructing the machine given an assumed model class, 'hierarchical machine reconstruction" describes a process of innovation to create a new model class. It detects regularities in a series of increasingly accurate

4258

A. Surana ct al.

models. The inductive jump to ;i higher computational level occurs by taking those regularities as the new representation. -machines reflect a balanced utilization of deterministic and random information processing, and this is discovered automatically during f-machine reconstruction. These machines are unique and optimal in the sense that they have maximal predictive power and minimum model size (hence satisfying Occam's Razor, i.e. causes should not be multiplied beyond necessity), e-machines provide a minimal description of the pattern or regularities in a system in the sense that ihe paltern is the algebraic structure determined by the causal states and their transitions. f-machines are also minimally stochastic. Hence, computational mechanics acts as a method for automatic pattern discovery. An ^:-machine is the organization of the process, or at least of the part of it which is relevant to our measurements. The p-machine that models the observed time series from a system can be used to define and calculate macroscopic or global properties that reflect the characteristic average information- processing capabilities of the system. Some of these include Entropy rate, E.xcess entropy and Statistical complexity fFeldman and Crutchfield 1998) and (Crutchfield and Feldman 2001). The entropy density indicates how predictable the system is. Excess entropy, on other hand, provides a measure of the apparent memory stored in a spatial configuration and represents how hard it is to predict. f>machine reconstruction leads to a natural measure of the statistical complexity of a process, namely the amount of information needed to specify the state of the ^-machine, i.e. the Shannon Entropy. Statistical complexity is distinct and dual from information theoretic entropies and dimension (Crutchfield and Young 1989). The existence of chaos shows that there is a rich variety of unpredictability that spans the two extremes: periodic and random behaviour. This behaviour between two extremes, while of intermediate information content, is more complex in that the most concise description (modelling) is an amalgam of regular and stochastic processes. An information theoretic description of this spectrum in terms of dynamical entropies measures raw diversity of temporal patterns. The dynamical entropies, however, do not measure directly the computational effort required in modelling the complex behaviour, which is what statistical complexity captures. Computational mechanics sets limits on how well processes can be predicted and shows how. at least in principle, those limits can be attained, ^-machines are what any prediction method would build, if only they could. Similar to ^-machine reconstruction, techniques exist which ean be used to discover casual architecture in memoryless transducers, transducers with memory and spatially extended systems (Shalizi and Crutchfield 2001). Computational mechanics can be used for modelling and prediction in supply chains in the following way: In systems like supply chains, it is difficult lo define analogues of various thermodynamic quantities like energy, temperature, pressure, etc. as we do for physical systems. Each component in the network has cognition, which is absent in physical systems such as a molecule of a gas. Because of such difliculties. statistical mechanics cannot be applied directly to build predietion models for supply chains. As discussed previously by not requiring a Hamiltonian (the energy-like function), computational mechanics is still applicable in the case of supply chains.

Supply-chain networks: a conip/e.x aclapiive systems perspective can he built to discover pallerns in the behaviour of various quantities in supply chains like the inventory levels, demand fluctuations, etc. e-machines can be used for prediction through a process known as "synchronization" (Crutchlield and Feldman 2003). f-machines can be used to calculate various global properties like entropy rate, excess entropy and statistical complexity that reflect how the system stores and processes information. The significance of these quantities has been discussed earlier. We can also quantify notions of Complexity. Emer^cmv and SclfOr^anization in terms of various information measures derived from ^-machines. By evaluating such quantities, we can compare the complexity of different supply chains and quantify the extent to which the network is showing emergence. We can also infer when a suppiy chain is undergoing self-organization and to what extent. Such quantification can help us to compare precisely what policies or cognitive capabilities possessed by individual agents can lead to different degrees of emergence and selforganization. Hence, we can decide to what extent we desire to enforce the control and to what extent we want to let the network emerge.

4259

7. Network dynamics The ubiquity of networks in the social, biological and physical sciences and in technology leads naturally to an important set of common problems., which are being currently studied under the rubric of "Network Dynamics' (Strogatz 2001). Structure always alVects function, and it is important to consider dynamical and structural complexity together in the study of networks. For instance, ihe topology of social networks affects the spread of information and disease, and the topology of the power grid aflects the robustness and stability of power transmission. The dilTercnt problem areas in network dynamics are discussed below. One area of research in this field has been primarily concerned with the dynamical complexity in regular networks without regard to other network topologies. While the collective behaviour depends on the details of the network, some generalizations can still be drawn (Strogatz 2001). For instance, if the dynamical system at each node has stable fixed points and no other attractor. the network tends to lock into a static fixed pattern. If the nodes have competing interactions, the network may display an enormous number oC locally stable equilibria. In the intermediate case where each node has a stable limit cycle, synchronization and patterns like travelling waves can be observed. For non-identical oscillators, the temporal analogue of phase transition can be seen with the control parameter as Ihe coupling coefficient. At the opposite extreme, if each node has an identical chaotic attractor., the network can synchronize their erratic fluctuations. For a wide range oT network topologies, synchronized chaos requires that the coupling be neither too weak nor too strong; otherwise, spatial instabilities are triggered. Related lines of research that address networks of identical chaotic maps are cotiplcd map lattiees (Kaneko and Tsuda 2000) and cellulur automutti (Wolfram 1994). However, these systems have

4260

A. Surana et al.

been used mainly as lestbeds for exploring spatio-temporal chaos und pattern formation in the simplest mathematical settings, rather ihan as models of real systems. The second area in network dynamics is concerned with characterizing the network structure. The network structure or topologies in general can vary from completely regular, like chains, grids, lattices and fully connected, to completely random. Moreover, the graphs can be directed or undirected and cyclic or acyclic. In order to characterize topological properties of the graphs, various statistical quantities have been defined. Most important of them include average path length, clustering coeRieient, degree distributions, size of giant component and various spectral properties. A review of the main models and analytical tools, covering regular graphs, random graphs, generalized random graphs, small-world and scale-free networks, as well as the interplay between topology and the network's robustness against failures and attacks ean be found in Albert {2000b). Albert and Barahasi (2002). Albert et a!. {2002). Callaway et al. (2000) and Dorogovtsev and Mendes (2002). Classic random graphs were introduced by Erdos and Renyi (Bollobas 1985) and have been the most thoroughly studied models of networks. Such graphs have a Poisson degree distribution and statistically uncorrelated vertices. At large ,V {total number of nodes in the graph) and large enough p (the probability that two arbitrary vertices are conneeted), a giant connected component appears in the network, a proeess known as percohuion. Random graphs exhibit u low average path length and a low clustering coeffieient. Regular networks, on other hand, show a high clustering coefiicient and also a greater average path length compared with the random graphs of similar size. The networks found in the real world, however, are neither eompletely regular nor completely random. Instead, we see 'small world' and 'scale free* characteristics for many real networks like social networks. Internet, WWW. power grids, collaboration networks, ecological and metabolic networks, to name a few. In order to describe the transition from a regular network to a random network. Watts and Strogatz introduced the so-called small-world graphs as models of social networks (Watts and Strogatz 1998) and (Newman 2000). This model exhibits a high degree of clustering, as in the regular network, and a small average distance between vertiees, as in the classic random graphs. A common feature of (his model with a random graph model is that the connectivity distribution of the network peaks at an average value and decays exponentially. Such an exponential network is homogeneous in nature: eaeh node has roughly the same number of connections. Beeause of the high degree of clustering, the models of dynamical systems with small-world coupling display an enhanced signal-propagation speed, rapid disease propagation. and synchronizability (Watts and Strogatz 199S. Newman 2002). Another significant recent discovery in the field of complex networks is that the connectivity distributions of a number of large-sealc and complex networks, including the WWW. Internet, and metabolic networks, satisfy the power law P{k) % k~'^, where P{k) is the probability that a node in the network is connected to A other nodes, and y is a positive real number (Albert et al. 2000a Barabasi e! al. 2000. Barabasi 2001). Since power-laws are free of the characteristic scale, networks that satisfy these laws are called scale-iree*. A seale-free network is inhomogeneous in nature: most nodes have a few connections, and a small but statistically significant number of nodes have many connections. The average path length is smaller in the

Supply-chain nclMorks: a complex adaptive sysWnis pcr.'ipeclive

4261

scale-free network than in a random graph, indicating that the heterogeneous scale-free topology is more eiFicient in bringing the nodes closer than homogenous topology of the random graphs. The clustering coefticient of the scale-free network is about five times higher than that of the random graph, and this factor slowly increases with the number of nodes. It has been shown that it is practically impossible to achieve synchronization in a nearest-neighbour coupled network (regular connectivity) if the network is sufficiently large. However, it is quite easy to achieve synchronization in a scale-free dynamical network no matter how large the network is (Wang and Chen 2002). Moreover, the synchronizability of a scale-free dynamical network is robust against random removal of nodes but is fragile to specific removal of the most highly connected nodes. The scale-free property and high degree of clustering (the small world effect) are not exclusive for a large number of real networks. Yet. most models proposed to describe the topology of complex networks have difficulty in capturing simultaneously these two features. It has been shown in Ravasz and Barabasi (2003) that these two features are the consequence of a hierarchical organization present in the networks. This argument also agrees with that proposed by Herbert Simon (Simon 1997), who argues: ... we could expect complex systems to be hierarchies in a world in which complexity has to evolve from simplicity. In their dynamics, hierarchies have a properly, near decomposability. that greatly simplities their behaviour. Near decomposabilily also simplifies the description of eomplex systems and makes it easier lo understand how ihe information needed for the development of ihe system ean be stored in reasonable compass. Indeed, many networks arc fundamentally modular: one can easily identify groups of nodes that arc highly interconnected with each other but have few or no links to nodes outside the group to which they belong. This clearly identifiable modular organization is at the origin of high degree of clustering coefficient. On the other hand, these modules can be organized in a hierarchical fashion into increasingly large groups, giving rise to 'hierarchical networks", while still maintaining the scale-free topology. Thus, modularity, scalc-frcc character and a high degree of clustering can be achieved under a common roof. Moreover, in hierarchical networks, the degree of clustering characterizing the different groups follows a strict scaling law, which can be used to identify the presence of hierarchical structure in real networks. The mathematical theory of graphs with arbitrary degree distributions known as generalized random graphs' can be found in Newman el nl. (2001) and Newman (2003). Using the 'generating function formulation", the authors have been able to solve the percolation problem (i.e. have found conditions for predicting the appearance of a giant component) and have obtained formulae for calculating the clustering coefficient and average path length for generalized random graphs. The authors have proposed and studied models of propagation of diseases, failures, fads and synchronization on such graphs and have extended their results for bipartite and directed graphs. Network dynamics, though in its infancy, promises a formal framework to characterize the organizational and functional aspects in supply chains (Thadakamalla ct al. 2004). With the changing trends in supply chains, many new issues have become critical: organizational resistance to change, inter-functional or inter-

4262

A. Suntiut ct al.

organizational conflicts, relationship management, and consumer and market behaviour. Such problems are ill structured, behavioural, and cannot be easily addressed by analytical tools such as mathematical programming. Successful supply-chain integration depends on the supply-chain partners' ability to synchronize and share real-time informalion. The establishment of collaborative relationship among supply-chain partners is a pre-requisite to information sharing. As a result, successful supply-ehain management relies on systematically studying questions like 1. What are the robust architectures for collaboration, and what arc the coordination strategies that lead to such architectures? 2. If different entities make decisions on whether or not to cooperate on the basis of imperfect information about the group activity, and incorporate expectations on how their decision will affect other entities, can overall cooperation be sustained for long periods of time? 3. How do the expectations, group size, and diversity affect coordination and cooperation? 4. Which kinds of organizations are most able to sustain ongoing collective action, and how might such organizations evolve over time? Network dynamics addresses many such questions and should be explored in the context of supply chains.

8. Conclusions and future work

The idea of managing the whole supply chain and transforming it into a highly autonomous, dynamic, agile, adaptive and reconfigurable network certainly provides an appealing vision for managers. The infrastructure provided by information technology has made this vision partially realizable. But the inherent complexity of supply chains makes the effieient utilization of information technology an elusive endeavour. Tackling this complexity has been beyond the existing tools and techniques and requires revival and extensions. As a result, we emphasized in this paper that in order to effectively understand a supply-chain network, it should be treated as a CAS. We laid down some initial ideas for the extension of modelling and analysis of supply ehains using the concepts, tools and techniques arising in the study of CAS. As future work, we need to verify the feasibility and usefulness of the proposed techniques in the context of large-scale supply chains.

Acknowledgements

The authors wish to acknowledge DARPA (Grant No. MDA972-I-I-0038 under the UltraLog Programme) tor their generous support for this researeh. In addition, the partial support provided by NSF (Grant No. DM1I-()O75584) to Professor Kumara is greatly appreciated. The authors wish to thank the anonymous reviewers for their comments and valuable suggestions.

Siippiy-ehain networks: a cnniptex adaptive systems perspective

4263

References
Abarbanel, H.D.I., Tiie .-inalysis of Observed Chaotic Data. 1996 (Springer: New York). Abarbanel. H.D.I, and Kennel, M-B,. Local false nearest neighbors and dynamical diniL'nsions from observed chaotic data. Phys. Rev. E. 199."*. 47. 3057 3068. Adami. C , Introduction to Artificial Life. I99S (Springer: New York). Albert. R, and Barabasi. A,L., Statisiical mechanics of complex networks. Rev. Mod. Plus.., 2002, 74, 47. Albert. R.. Barabasi, A.L.. Jcong, H. and Bianconi, CJ.. Power-law distribiilion of the World Wide Web. Science. 2000, 287. 2115. Albert R.. Jcong. H.. Barabasi. A.L.. Error and attack tolerance of complex networks. Nature. 2000. 406. 378 3S2. Balakrishnan. A.. Kumara. S. and Snnd;iresan, S.. Exploiting information technologies for product realization. Inform. Syst. Front. J. Res. Innov.. 1999, 1(1). 25 50. Barabasi. A.L.. The physics of web. Phys. World. July 2001. Barabasi, A.L.. Albert. R. and Jeong. H.. Scale-free characteristics ol" random networks: The topology of the World Wide Web. Phy.s-ica A. 2000, 281, 69-77. Baranger, M.. Chaos, complexity, and entropy: a physics talk for non-physicists. Available online at: http://necsi.org. projccts/barangcrccc.pdf (accessed May 2005). Bar-Yam, Y,, Dynamics of Complex Systems. 1997 (Addison-Wcsley: Reading, MA). Bollobas. B,. Random Graphs, 1985 (Academic Press: London). Callaway. D.S.. Newman, M.E.J.. Strogatz. S.H, and Wiitts, D.J.. Network robustness and fragility: Percolation on random graphs. Phy.s. Rev. Lett.. 2000, 85, 546S 5471. Carlson. J.M., Doyle, J., Highly optimised tolerance: a mechanism for power laws in designed systems. Pins. Rev. E. 1999, 60(2), 1412 1427. Casdalgi. M., Nonlinear prediction of chaotic time scries. Physica D. 1989, 35. 335-356. Choi, T.Y., Dooley, K.J., Ruanglusanathan. M,, Supply networks and complex adaptive systems: eontrol versus emergence. J. Operat. Manage.. 2001, 19(3), 351 366. Cooper, M.C\. Lambert. D.M. and Pagh. J.D.. Supply chain management: more than a new name lor logistics. ////. J. Logist. Manai^e.. 1997. 8(1), 1-13. CrutcliHekl, J-P-, Knowledge and meaning . . . chaos and complexity. In Modeling Complex Systems, edited by L. Lam and H.C. Morris, 1992 (Springer: Berlin), pp. 66 101. Crutchfield, J.P., The calculi of emergence: computation, dynamics and induction. Phvsica D. 1994, 75. 11-54. Crutchfield, J.P. and Young, K.. Inferring statistical complexity. Phys. Rev. Lett.. 1989, 63, 105 108. Crutchtield, J.P. and hcldmiin, D.P., Synchronizing to the environment: information theoretic constraints on agent learning. Adv. Complex Sy.st.. 2001, 4, 251 264. Crutchtield. J.P, and Fcldman. D.P.. Regularities unseen, randomness observed: levels of entropy convergence. Chaos. 2003, 13, 25 54. Csete. M.E. and Doyle, J., Reverse engineering of biological complexity. Seienee. 2002. 295, 1664. Dorogovtsev. S.N. and Mendes, J.F.F.. Evolution of networks. Adv. Phys.. 2002. 51. 1079-1187. Erramilli, A. and Forys, L.J., Oscillations and chaos in a flow model of a switching system.
lEI^EJ. Select. Areas Commun.. 1991, 9(2), 171-178.

Farmer, J.D., Ott. F. and Yorke, J.A., The dimension of chaotic attractors. Physiea D, 1983, 7. 153-180. Farmer, J.D. and Sidorowich, J.J.. Predicting chaotic time-series. Phvs. Rev. Lett.. 1987, 59(8), 845- 848. Feichlinger. G.. Hommes, C H . and Herold, W.. Chaos in a simple deterministic queuing system. ZOR- Math. Meth. Oper. Res.. 1994, 40, 109 119. Feldman, D.P. and Crutchfield, J.P., Discovering non-critical organization: statistical mechanical, information theorclic and computational \ iews of patterns in one-dimensional spin systems. Santa Fe Institute Working Paper 98 04 026, 1998. Flake, G.W., The Compntationul Beauty of Nature. 1998 (MIT Press: Cambridge, MA), Forrester, J.W.. Industrial Dynamics. 1961 (MIT Press: Cambridge, MA).

4264

A. Siircma ct al.

Fraser, A.M. und Swinney. H.L.. Independent eoordinates for strange uttractors from mutual information. Plus. Rev. A. I9H3. 33(2). 1134 1140. Ghosh, S., The role of modeling and asynehronous distributed simulation in analyzing complex sysiems of the future. Inform. Sy.st. Front. J. Res. Iniiov., 2002. 4(2). 166 171. Glance. N.S., Dynamics with expectations. PhD thesis. Physics Department, Stanford University. 1993. Hogg, T. and Hiiberman, B.A., The behavior of computational ecologies. In The Ecology of Compuiaiioiu edited by B.A. Htiberman, pp. 77 116, I9SS (Elsevier Science: Amsterdam). Hogg, T. and Huberman. B.A.. Controlling chaos in dislribuled systems. IEEE Trans, on Syslcm.-i. Man and Cybernetics. 1991, 21, 1325-1332. Katieko, K. and Tsuda, I., Complex Systems; Chaos and BeyondA constructive approach wilh applications in life .sciences. 2000 (Springer: Berlin). Kennel, M.. Biown, R. and Abarbanel, H.D.I.. Determining embedding dimension for phase-space reconstruction using a geometrical construction. Pins. Rev. A. 1992, 45(6), 3403-306S. Kephart. J.O.. Hogg. T. and Huberman, B.A., Dynamics of computational eeosystems. Phys. Rev. A, I9ij9, 40(1), 404-421. Kephart, J.O., Hogg. T. and Huberman. B.A., Collective behavior o\' predictive agents. Phy.sica D, 1990.42,48 65. Kumara, S.. Ranjan. P.. Surana, A. and Narayanan, V.. Decision making in logistics: A chaos theory based analysis. Ann. Int. Insl. Prod. En^. Res. (Ann. CIRP), 2003, 1, } ^ \ 384. Lee. S.. Gautani, N., Kumara, S.. Hong, Y., Gupta, H., Surana, A.. Narayanan, V., Thadakamalla, H., Brinn. M. and Greaves, M., Situation identifieation using dynamic parameters in eomplex agent-based planning systems. Inlelt. Eng. Sy.st. Artif Neural Nelm>rks, 2002. 12. 555- 560. Lind, D. and Marcus. B.. An introduction to symbolic dynamics and coding, 1995 (Cambridge University Press: New York). Llyod. S. and Slotine. J.J.E., Information Uieorelic tools for stable adaptation and learning. Int. J. Adapt. Control Signal Process., 1996, 10, 499-530. Maxion, R-A., Toward diagnosis as an emergent behavior in a network ecosystem. Physica D. 1990,42,66-84. Min, H. and Zhou, G.. Supply chain modelinc: past, present and future. Coniput. Ind. Eng.. 2002,43, 231 249. Mukherjee, S.. Osuna, H. and Girosi, F., Nonlinear prediction of chaotic time series using support vector machines. In IEEE Work.shop on Neural Networks for Signal Proeessing III. 1997, pp. 511-519. Newman, M.H.J., Models of the small world. J. Stat. Phys.. 2000, 101, 819 841. Newman. M.F.J., The spread of epidemic disease on networks. Phys. Rev. E, 2002, 66. Newman. M.F.J., Random graphs as models of networks. In Handbook of Graphs and Networks, edited by S. Bornholdt and H.G. Schuster, 2003 (Wiley-VCH^ Berlin). Newman, M.F.J., Strogatz, S-H. and Watts, D.J., Random graphs with arbitrary degree distribution and their applications. Phys. Rev. E, 2001, 64. Ott, E., Chaos in Dynamical Sy.stems. 1996 (Cambridge University Press: Cambridge). Powell, M.J.D., Radial basis function approximation to polynomials. Preprint University of Cambridge, 1987. Rasmussen, D.R. and Moseklide, M., Bifurcations and chaos in generie management model. Eur. ./. Oper. Res.. 1988, 35, 80 88. Ravasz, E. and Barabasi. A.L.. Hierarchical organization in complex networks. Plus. Rev. E. 2003. 67. Sano, M. and Sawada, Y., Measurement of the Lyapunov Spectrum form a chaotic time series. Phys. Rev. Lett.. 1985, 55, 1082-1084. ' Sawhili, B.K., Self-organised criticality and complexity theory. In !993 Lectures in Complex Systems, edited by L. Nadel and D.L. Stein, pp. 143 170. 1995 (Addison-Wesley: Reading, MA).

Supplx-tiutiii iielni>rk.\: a coiupte.x adaptive .vr.v/c/j/.v pvrspective

4265

Schieritz, N. and Grobler. A.. Emcfgcnl structures in supply chains A study integriiting agenl-bascd and system dynamics modeling, in 36th Annual Hawaii International Conference on Svsteni Sciemes. Big Inland. HI. 2003. Shalizi, C.R. and Crutchlield. .I.P.. fomptitiilional meciianics; pattern and prediction, Mruciure and simplicity. ./. Stat. Pliy.s., 2001. 104. 816-879. Shali/i. C R . . Causal architecture, complexity and self-organizaiion in time series and cellular automata. Available online at: http://www.santare.edu--shalizi, thesis, 2005 (accessed May 2005). Simon, H.A.. The Sciences ol the .Artificial. 3rd ed.. \997 (The MIT Press. Cambridge. MA). Strogatz, S.H., Nonlinear Drnaniics and Chaos, 1994 (Addison-Wesley: Rending. MA). Strogatz, S.H., Exploring complex networks. Nature. 2001. 410. 268 276. Takens, F., Deteeting Strange Attraetor in Turbulence. In LS. Young, Editor. Dynamical Systems and Turbulence. Led we Notes in Matlwmaiiis. 1981, 898, 366-381. (Springer, New York). Thadakamaila. H.P.. Raghavan, U.N.. Kumara, S. and Albert, R., Stir\ivability of multiagenl-bascd stipply networks: a topological perspecti\e. IEEE Intell. Svst.. 2004, 19(5), 24-31. Wang. X.F. and Chen. G., Synchronization in scale-free dynamical networks: robustness and fragility. /EKE Trans. Circuits and Systems I Eundant- Thcorv Applic. 2002, 49( 1), 54 62. Watts. D.J. and Strogatz. S.H.. Collective dynamics of small-uorld" neuvorks. Nature. 1998, 393. 440 442, Wolfram, S.. Cellular Automata and Complexity: Collected Papers, 1994 (Addison-Wesley: Readinu. MA).