دو صفحه اول مقاله شما، جهت اثبات حسن نیت پرتال ارسال گردید- برای دریافت متن کامل مقاله خود از سرویس دانلود مقالات علمی استفاده کنید و یا با لینک کردن ما در سایت و یا وبلاگ خود، آن را به رایگان دریافت کنید.

 

A multi-agent based framework for supply chain risk management

• Mihalis Giannakis^{a, ,} ,
• Michalis Louis^b

• ^a Operations Management Group, Warwick Business School, The University of Warwick, Coventry CV4 7AL, UK
• ^b The University of Warwick, Coventry CV4 7AL, UK

• Received 10 September 2009. Revised 30 April 2010. Accepted 5 May 2010. Available online 14 May 2010.

• http://dx.doi.org/10.1016/j.pursup.2010.05.001, How to Cite or Link Using DOI
• Cited by in Scopus (7)

• Permissions & Reprints

---

Keywords

• Multi-agent systems;
• Supply chain management;
• Risk management

---

1. Introduction

The increasing call for mass customization in many industries has made today’s global supply chains very complex, requiring a multitude of parallel information and physical flows to be controlled to ensure high customer service levels. This increased complexity raises the level of uncertainty and risks that companies are faced with Manuj and Mentzer (2008). The wide range of risks along the supply chain (both from supply and demand side) may impose negative implications upon supply chain performance. There is an eminent need for organizations to have necessary strategies to manage these risks and disruptions, so that they can achieve the necessary level of agility for effective mass customization.

Constructive collaboration among business partners in supply chains is vital in any attempt to mitigate risks and ameliorate disruptions, to achieve responsiveness and to offer a high customer service level (Hallikas et al., 2004). Many successful modern organizations have shifted from an opportunistic dogma of cooperation to a synergetic ethos of collaboration and aligned their supply chain processes.

The use of Information and Communication Technology (ICT) tools is perceived as a paramount facilitator for the realisation of this collaborative perception, offering the capabilities of information sharing, customer sensitivity and process integration (Wu and Angelis, 2007). The conventional IT, however, (which is based on legacy systems) has not provided sustainable solutions for collaborative Supply Chain Management (SCM). It lacks real-time adaptability in supply chains and focuses on dyadic contexts of collaboration rather than collaboration amongst a plethora of partners (Akkermans et al., 2003). It is also characterized by inflexibility for the reconfiguration of supply chains processes and high development and maintenance costs (Botta–Genoulaz et al., 2005).

The use of multi-agent modelling (a sub category of artificial intelligence) can be an alternative decision making tool for collaboration within supply chains. In computer science, an agent can be defined as a software entity, which is autonomous to accomplish its design objectives, considered as a part of an overall objective, through the axiom of communication and coordination with other agents (Gilbert, 2007). Through this paradigm of software architecture, the management of supply chain processes can be perceived as facilitated by several autonomous decision making entities (software agents), each responsible for specific activities and performing different roles. These agents interact and cooperate with other agents, within and across organizations, in order to solve problems beyond their individual knowledge or expertise, and to promote a higher performance for the entire system (Stone and Veloso, 2000).

In this paper, a multi-agent based framework is proposed as the conceptual basis for the design of a DSS that facilitates collaborative disruption risk management in manufacturing supply chains. The framework supports the fulfilment of production, event and disruption risk management constituted by coordination, communication and task agents and draws on principles and theories of SCM, agent based simulation and computer science.

The remainder of the paper is organised in four sections. In the first section, the usefulness of a multi-agent system (MAS) framework for supply chain risk management (SCRM) is discussed through a brief review of an expansive SCRM literature, a comparison between conventional IT solutions and MAS and a discussion of the application of software agents to different supply chain problems. The second section presents the analytical approach that has been utilized, the process for the development of the framework and its features in detail. With the use of a hypothesized scenario, the third section presents the processes for supply chain disruption management that an MAS designed with the logical structure of the proposed framework will follow. The paper concludes with a discussion of the limitations and managerial implications of the framework and potential extension of the research.

2. Literature review

The SCM literature is rife of studies that investigate supply chain risk phenomena and provide models for the analysis and mitigation of several types of supply chain risks (financial, operational and strategic) that occur both in the supply and demand side of supply chains. Zsidisin and Ritchie (2008) and Juttner (2005) provide comprehensive reviews of the literature and models used for an effective SCRM. Several authors have also proposed disruption management strategies for SCRM. Zsidisin and Smith (2005) discuss how early supplier involvement can reduce the likelihood of supply disruptions and Chopra and Sodhi (2004) highlight mitigation strategies for different types of risks, which manufacturing organizations apply to deal with uncertainty. They identify drivers for a wide variety of different risks and pinpoint alternative proactive mitigation strategies for each corresponding risk. Other researchers have focused specifically on information systems risks. Kleindorfer and Saad (2005) underline the importance of collaborative information sharing as the vehicle to shed light upon vulnerabilities within the supply chain, to manage disruptions under a cost efficient manner and to devise strategies for their effective control in a crisis situation.

The existing studies that are found in the literature have been instrumental in identifying and analysing several causes of disruptions and risks in supply chains. The majority of them, however, are based on case studies and empirical evidence, which limits their generalizability beyond the specific contexts and the extension of the strategy beyond dyadic relationships. Melnyk et al. (2009) provide an innovative formalized methodology in mitigating disruptions in supply chains with the use of discrete event simulation. Their approach extends previous studies; however, their study is limited by the lack of real-time adaptability of their model that is evidenced in supply chains. Methodologies that can overcome this limitation like agent based modelling are considered very useful in this respect.

The agent based technology is acknowledged as one of the most promising technologies for effective management of supply chains that are characterized by high levels of uncertainty ( [Brooks and Davenport, 2004] and [Lou et al., 2004]) due to its vital characteristics of:

•

Autonomy: agents are aware of their environment operating without human intervention to some extent in order to fulfil their objectives (Jennings and Wooldridge, 1995)

•

Social ability: an agent can interact with other agents or humans through the use of an agent communication language (Moyaux and Chaib-draa, 2006).

•

Reactivity: agents can perceive their environment and respond to specific changes in this environment (Parunak, 1999)

•

Pro-activeness: agents do not simply respond to changes in their environment, but can initiate actions (Moyaux and Chaib-draa, 2006).

Despite the fact that the technology is relatively old and well known, in today’s business reality multi-agent software platforms are not used widely for the management of business processes. Electronic resource planning systems (ERP) are still the most common software selection for managing a company’s processes (Burka et al., 2005). The latest advancements of ERP systems that incorporate the integration of SCM processes give rise to promises that they can lead to significant improvements (Kelle and Akbulut, 2005).

The efficacy of conventional ERP systems, however, to provide real time synchronization among supply chain partners – which is necessary for effective SCRM – is limited. Their design has been based on the need for integration with legacy systems. This restricts their ability to meet the high level of agility required for decentralized control (Karwowski et al., 2007). The inflexible nature of ERP systems to adapt to rapidly changing business environments in a cost efficient manner, has forced many organizations to fit their processes according to the capabilities of the ERP system in order to save funds and time (Helo et al., 2005).

The latest generation of ERP systems that utilize Service Oriented Architecture and web services for process management has the potential of providing real time information on several activities in an enterprise (Tarantilis et al., 2008), which can speed up the decision making process and lead to better integration amongst supply chain partners. Their modular design increases their flexibility, making them more applicable to several managerial processes. Their computational capability for managing very complex processes, however, is limited (Helo et al., 2005), and their role is confined only in the timely flow of information about the condition of several processes. A considerable amount of time and funds is also required in order to yield benefits from the latest ERP technologies, implying that only organizations that can afford the required high level of investment will be able to achieve responsiveness, and overall to reinforce their competitive advantage. The multi-agent architecture for an SCM (Fig. 1) offers an opportunity to overcome the shortcomings of conventional ERP technology, providing a plethora of advantages.

Through their learning capability, an MAS can demonstrate efficiently the proactive and autonomous behaviour of the participating agents in mitigating risks and rectifying supply chain disruptions in real time ( [Kwon et al., 2007] and [Lu and Wang, 2007]). They can also promote a high level of cross organizational collaboration in a computational and cost efficient manner (Swaminathan et al., 1998).

The inherent distributed nature of agent based technology (in that, a problem solution is distributed to different agents) gives the significant advantage of ease in dealing with the high level of complexity of global supply chains in contrast with conventional information technology (Akkermans et al., 2003). This is enhanced by the fact that each of the agents has a specific expertise and a computational efficiency in dealing with this complexity in combination with easiness of development in a short time frame (Lu and Wang, 2007). With this approach, re-configuration of the whole supply chain can become a reality in a timely fashion and with a low cost.

These benefits can be materialised by incorporating existing legacy systems. An expert system for inventory management, or an ERP system, for example, can be integrated with multi-agent software. The main technique applied is to “wrap” around the legacy code with an agent based software, without the need to rewrite the whole legacy code, in order to “agentify” it into a normal agent (Genesereth and Ketchpel, 1994). On this basis, the wrapping software is used as a “facilitator” to interpret messages from an agent to legacy software and vice versa, aiming at providing an understandable communication platform for both sides (Davidsson and Wernstedt, 2002).

MAS have been used in modelling a multitude of supply chain phenomena: for the identification of negotiation problems in supply chains (Chen et al., 2004), in production and control processes (Caridi and Cavalieri, 2004) in distribution (Swaminathan et al., 1998), and for inventory and demand forecasting (for a thorough literature review of an MAS application on an SCM see Beamon, 1998). Agent based technology has also been utilised for the management of disruptions within a supply chain in some studies. Kimbrough et al. (2002), for example, use it for the reduction of the bullwhip effect through modelling a supply chain with agents. Kwon et al. (2007) propose an agent based framework based on experimental design that deals with supply and demand uncertainty and Bansal et al. (2005) provide a generalized collaborative framework for disruption management oriented to the reality of refinery supply chains.

The papers that deal with the application of an MAS in disruption management focus on particular supply chain risks and/or contexts, but do not explore the learning process of the agent based models to ameliorate abnormalities in supply chain processes and to integrate decisions across the supply chain. Their approach to SCRM is limited as they do not provide the basis for an integrated SCM framework for decision making.

In this paper, these previous works are synthesised and extended through the development of a framework for supply chain disruption management that deals with amelioration of abnormalities in manufacturing supply chains that incorporates basic risk management theory. The proposed framework is enriched with case-based reasoning (decisions for previous cases can be utilized as instances for current decision making), in order to enable the learning capability of the agent (Leak, 1994), by taking into consideration integration issues, so that it can be used as an add-on module to legacy systems.

3. Framework for a disruption management system

For the development of the framework, an analytical sequential approach was adopted. First the organizational design of the framework was formulated and embedded in an overarching agent based SCM model (an artificial society constituted by software agents). Then the roles for each of the agents within the disruption management framework were defined and a detailed description of the responsibilities for each of these roles was provided. The interactions among these agents were subsequently modelled by analyzing several risk identification and mitigation processes.

In Fig. 2, the basic components of a generic multi-agent based model for an SCM are presented. This model is constituted by three modules: (a) production fulfilment processes (e.g. order management, manufacturing, procurement, logistics); (b) supply chain event management; and (c) disruption risk management processes.

The production fulfilment module coordinates the activities of different supply chain partners for the fulfilment of orders through the supply, production and delivery processes. The supply chain event management module is responsible for monitoring the actual fulfilment of specific orders along the supply chain. The role of the disruption risk management module is to initiate the necessary coordination among the agents, when a risk through a potential disruption is identified, related to a specific order or to the overall operational performance.

The focus of this paper is on the disruption risk management module. This module is initiated by the monitoring process of the MAS and incorporates several sequential processes that are described in this section. It synthesizes the monitoring of supply chain events with the disruption management processes. It is comprised of five software agents:

•

Communication agent: responsible for the effective communication between the agents of supply chain partners. In common terms, it facilitates the flow of information among the partners of a supply chain. For example, a logistics service provider (LSP) and a manufacturer can communicate through the facilitation of this agent.

•

Coordination agent: responsible for coordinating the agents of the module within the limits of an organization. The coordination agent of a focal company (e.g. an OEM) orchestrates the disruption management process only for the processes of this company.

•

Monitoring agent: responsible for providing the required monitoring information, by gathering and analyzing corresponding data from all the collaborating parties. It has the ability to trigger an alarm, when an abnormal situation occurs. In effect, it provides inter-organizational visibility regarding either the normality or abnormality of processes related to the fulfilment of an order. For instance, from a supplier’s perspective, it provides supervision of the production process, by monitoring the actual production and comparing it to the production schedule.

•

Wrapper agents: these agents can offer information integration among legacy and agent software. For example, an expert system for inventory management or an ERP system can be integrated with agent software. The main technique applied is to “wrap” around the legacy code with agent software, so that to “agentify” it into a normal agent. On this basis, the wrapping software is used as a “facilitator” for the interpretation of messages from agent to legacy software and vice versa, aiming to provide an understandable communication for both sides (Davidsson and Wernstedt, 2002).

•

Disruption manager agent: it is the main cell of the disruption management module. It is constituted by a rectification proposer (a software agent responsible for the suggestion of corrective actions), a built-in simulator that accounts for the risk assessment and the optimization processes, and a learning component that leverages successful past decisions as cases for future use in similar situations.

The competences and activity limits of the aforementioned agents are summarized in Table 1.

The disruption risk management process is triggered when an unusual event is identified during the order fulfilment process (Fig. 2). The process can be described as follows: after the acceptance by the customer of the proposed delivery time and price, the coordination agent triggers a monitoring process for the specific order. This entails the creation of a monitoring agent devoted to this order (step 1). Consecutively, in a cascading pattern, monitoring agents are generated by the local coordination agents of the supply chain partners, which are directly linked with the specific order (first and second tier suppliers). In this respect, for each supply chain partner that is related to the specific order, a monitoring agent is triggered by the local coordination agent. A recursive process of information gathering then begins (step 2). In the case that an abnormal event is identified within this process, as a cause of a potential disruption, the disruption risk management module is activated (step 3).

Assume, for example, that an abnormal event is identified in the manufacturing process of the second tier supplier by the corresponding monitoring agent. This may be an unusual fluctuation in the manufacturing process, either due to a labor strike, or a problem with manufacturing equipment. The monitoring agent that has identified the abnormal event alerts the local disruption risk management agent, so that the necessary guidelines will be provided, for the proactive mitigation of the situation. The disruption risk management process has begun. This process is illustrated in Fig. 3. A detailed description of the process for risk management and mitigation is presented in the next section.

3.1. Risk management process by the disruption manager

The risk management process is executed in four stages: risk identification, risk assessment, decision and implementation of risk management actions and optimization (Hallikas et al., 2004). The process involves collaboration amongst supply chain partners through exchange of information and allocation of specific roles, in order to enable mutual risk management.

3.1.1. Risk identification

This is the fundamental stage of the entire risk management process, where risks are identified through the use of quantitative models. The backbone of this process is based upon the monitoring of various key performance indicators (KPIs) related to the performance of supply chain partners (e.g. suppliers, LSP). The level of an in-stock inventory, production throughput, capacity utilization and delivery lead times are some of the KPIs that can be used to identify an abnormal situation that may involve a potential risk.

The actual values of those KPIs are monitored within a specific time frame and compared with predefined values that are described either in an agreement among partners, or unofficially when the type of relationship does not require an agreement. Statistical tests can identify significant deviations between the actual and the pre-defined values (Fig. 4). In case a significant deviation is identified, an alarm is triggered by the monitoring agent.

3.1.2. Risk assessment

This stage is necessary for the selection of suitable corrective management actions for the risk factors identified in the previous stage. The risk analysis takes into consideration a wide range of criteria such as the probability of occurrence of the event, the risk level and risk impact, and it prioritizes the risks according to the outcome of this process.

In conjunction with the outcome of the risk assessment, a description for the level of the impact using ordinal scales (e.g. no impact, minor impact, medium impact or serious impact) and the level of probability for the occurrence of the event (e.g. very unlikely, improbable event, moderate event, probable event, very probable ) can be given. This process is executed by a root cause identifier software, which is incorporated in the learning module of the disruption management agent. Through the monitoring of crucial KPI’s (e.g. delivery time, production output), the potential causes of the triggered alarm can be identified. For instance, in case of a significant delivery delay, the root cause identifier will initiate a process to trace the cause of this delay. In the case of inability on behalf of the LSP to deliver the order, the root cause identifier will “label” this incident as probable delays risk.

Subsequently, each of the potential remedies for any identified cause (that is about to become a risk), is evaluated using a built-in simulator. The success of a past decision applied to a specific risk is taken into consideration during the evaluation process. The best past responses to risks are selected, through the existence of a “risk portfolio”, improving the quality of the solutions. This utilization of previous successful knowledge to similar future situations is accomplished through the use of case-based reasoning (Leake, 1994). This system-learning methodology can be applied to the proposed framework through the structure depicted in Table 2.

For instance, in case an alarm is triggered at the side of a supplier due to the identification of an abnormal KPI, all the potential risks that arise (e.g. potential delays) are identified by the root identifier in the field “type of risk”. Potential remedies are then evaluated by the built-in simulator initially in terms of their feasibility and based on specific constraints (e.g. contractual agreements). Then, the expected costs of those feasible scenarios are calculated, generating a list of remedies for the risk that emerges along with the associated costs.

The mechanism that the built-in simulator adopts during the risk assessment process can be described using the model of Shavell (1984) for risk mitigation. With this model the level of the danger, the cost for its mitigation and the potential financial loss that leads to a disruption, are quantified. The following calculations are conducted hierarchically: (a) the probability P(y) for the disruption to become reality is estimated through the use of Failure Mode and Effect Analysis (FMEA) and/or formal mathematical models that can utilise linear regression, time series regression models and stochastic models (Dani, 2009); (b) the amount of financial loss L(y) for the specific disruption to become reality; (c) the investment cost y to mitigate the probability P(y), in order to lessen the specific disruption risk; (d) an optimal y^⁎ investment cost for the mitigation of the risk, in order to minimize the expected cost that is to arise in case the risk becomes a reality (Fig. 5).

3.1.3. Decision and implementation of risk management actions