Designing Agent based Models for Supply Networks

560 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
560
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
16
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Designing Agent based Models for Supply Networks

  1. 1. Agent Based Models for Supply Networks (extended abstract) Yifeng Zhang Sid Bhattacharyya (yzhang14@uic.edu, sidb@uic.edu) Information and Decision Sciences University of Illinois at Chicago May 24, 2010 Submitted to: Workshop on Revolutionary Strategies and Tactics in Research Design And Data Collection for E-Business Management Research International Conference on Electronic Commerce 2003.
  2. 2. 1 Introduction Advancements in information technology impact supply chain management in two ways: electronic integration and electronic markets (Malone, Yates et al. 1987). Electronic integration effect refers to strengthening existing relationships among partners of a supply chain. Electronic market effect refers to a shift towards decentralization that emphasizes dynamic and market-like relationships among partners of a supply chain. Both effects have been noticed in the real world. For instance, Vendor Managed Inventory and private B2B e-markets are examples of electronic integration effect. And public B2B e-markets (exchanges) are examples of electronic market effect. Malone et al. (1987) argued that market effect will dominate integration effect in the long run. This argument is consistent with transaction cost theory (Williamson 1975), which states that market structure carries higher transaction cost than hierarchical structure. As advances in information technology lowers transaction costs, the disadvantages of a market structure become less significant, and the structure will be more prevalent in supply chains (Croom 2001). The management challenge of supply chains (integrated) arises from various uncertainties, such as demand fluctuation, information distortion, and capacity constraints. The dynamic customer-vendor relationship in decentralized supply chain adds another component to these uncertainties. In traditional and electronically integrated supply chains, static long-term customer-vendor relationships are emphasized. The stable relationships allow a chain leader to perform system-wide coordination so as to control the uncertainties. In decentralized supply chains, the role of a chain leader diminishes, thus supply chain dynamics (Forrester 1961) result from dynamic interactions of participants. Little is known about the impact of this new type of dynamics. This study takes a step in this direction by examining impacts of marketization on several aspects of supply chain dynamics, including bullwhip effect (Lee, Padmanabhan et al. 1997) and service level. The widely observed bullwhip effect refers to the phenomenon that order variance increases up along a supply chain. The bullwhip effect has been attributed to demand signal processing, rationing, order batching, price variation, demand seasonality, and forecast error (Lee, Padmanabhan et al. 1997; Metters 1997). Service level is a major performance indicator in supply chains (Beamon and Chen 2001). It typically refers to the portion of customer demand that is fulfilled on time and measures the effectiveness of a supply chain. The impact of marketization was examined through comparisons of bullwhip effect and service level between networked (marketized) and divergent (integrated) supply chains. In divergent supply chains, each player has one and only one upstream partner, but can have multiple downstream partners (Beamon and Chen 2001). In networked supply chains, customers have choices on which vendor to use for each order they place. In divergent supply chains, customers always place orders to a same vendor. Two networked supply chain models are constructed, which differ in vendor selection
  3. 3. methods. In one model (network-random, NR) that is used as a baseline, customers select vendors randomly. In the other model (network-inventory, NI), customers select vendors who have the highest net inventory levels at the time of placing an order. Two models are considered for the divergent structure too. One model (divergent-single, DS) contains a single supply chain and the other (divergent-multiple, DM) contains multiple supply chains. These two divergent models are in consideration of the two different levels at which marketization may occur. At the industry sector level, marketization blurs lines between different supply chains. In traditional and electronically integrated supply chains, static long-term customer-vendor relationships are emphasized. Firms of a same industry sector are divided into different supply “chains”. Although there are interactions between different chains (Fransoo and Wouters 2000), the separation is so profound that each chain is often managed and analyzed independently. At the supply chain level, a single chain might break down into a network when the original chain leader is replaced by multiple peer players (Croom 2001). 2 Agent Based Modeling Agent-based modeling (ABM) is employed in this study. ABMs undertake a bottom-up approach to modeling of individual agents and the way they act and interact; the overall dynamics emerge from the collective interaction of heterogeneous agents. ABM is noted to be useful for problems that can’t be mathematically modeled, or where analytical solutions are not obtainable (Axtell 2000). Compared to equation-based modeling and simulation, ABM is more suitable for complex problems that have high degree of localization and distribution (Parunak, Savit et al. 1998). The Swarm1 agent based modeling toolkit is used. The supply network is modeled as a composite of a collection of retailers and vendors, together with outside customers and suppliers. Each agent in the supply network is defined by a number of attributes and actions that it takes, and the model incorporates various delays, informational and physical, in its operation. So the supply chain itself is defined. ABM can accommodate a much higher degree of detail and complexity than other approaches. This advantage makes it an ideal tool for leveraging detailed real-world data in studies of e-business phenomena. An agent-based model can be calibrated with real-time data and provides an opportunity to experiment with "reality" in a way that is difficult to do otherwise. ABMs can be aligned with observed behavior and phenomena at varying levels. Models can be ‘docked’ with analytical results by constraining the model to match assumptions and constraints in the analytical model. Individual agent models can also be set up to correspond to ‘rules’ followed by real-world supply network entities. Models can be calibrated and validated using data at the agent level, or at aggregated-agent levels as in, say, a vendor-group or at the level of the organization. 1 www.swarm.org
  4. 4. 3 The Models There are four types of agents used in this study. At the beginning of each period, a customer agent generates demands and pass them to buyer agents (demands follow a uniform distribution in the range of [10 20]). Buyer agents then perform three tasks in the following order: accept inbound goods (if any), dispatch outbound goods (if any), and place new orders (to vendor agents). Vendor agents then perform the same three tasks, but with slightly different rules, in the same order. And finally, a supplier agent fulfills vendor agents’ orders. Each model in the results reported here consists of one customer agent, ten buyer agents, and one supplier agent. The number of vendor agents differ among models. There are three vendor agents in model NR, NI and DM, but only one in model DS. All buyer agents form a buyer stage, and similarly all vendor agents form a vendor stage. Specific configurations of the models is shown in Figure 1. In the current setup, both buyers and vendors adopt a modified order-up-to-inventory replenishment policy. Order sizes are determined by the following two equations: O = U – I – P + (D – S) (1) U = (1 + IL + PL) * DMA + α * (1 + IL + PL) * DMV (2) Here, α: constant coefficient D: demand of current period DMA: moving average of demand DMV: moving variance of demand I: inventory at hand IL: information lead-time O: order size P: orders in pipeline (orders placed but not yet received) PL: physical lead-time S: shipment of current period U: order-up-to level Figure 1: Configurations of Models a: configuration of models NR and NI
  5. 5. Buyer 1 Vendor 1 Buyer 2 Buyer 3 Buyer 4 Supplier Vendor 2 Customer Vendor 3 Buyer 10 b: configuration of model DS Buyer 1 Buyer 2 Buyer 3 Buyer 4 Supplier Vendor 1 Buyer 5 Customer Buyer 6 Buyer 7 Buyer 10 c: configuration of model DM
  6. 6. Buyer 1 Vendor 1 Buyer 2 Buyer 3 Buyer 4 Supplier Vendor 2 Buyer 5 Customer Buyer 6 Buyer 7 Vendor 3 Buyer 10 4 Initial Results For the results given here, each model was operated for 200 periods, and use zero values for information delays (time for transmitting orders from originator to destination) (Chen 1999) (The models can take other values though). Orders are processed in a first-come- first-serve basis together with a no-partial-shipment policy2. Shipping of a period stops when all orders are fulfilled or inventory is insufficient to fulfill any order in queue. Once an order is shipped, it is received by the originator after a physical delay (Chen 1999), which is modeled as a uniformly distributed random variable in a specifiable range ([0 3] for experiments presented here, unless otherwise noted). Buyers and vendors handle inventory shortage differently. A buyer loses an order permanently if he can’t immediately (in the same period that the order is received) fulfill it. However, a vendor maintains orders that are not immediately fulfilled and tries to fulfill them in following periods. To alleviate effects of randomness, simulations for each model were replicated 30 times, each time with a different set of random seeds. A separate random number generator was used by the customer agent so that the demands remain identical across models. Initial inventory levels are set for supply stages to increase comparability among models. Prior to experimental runs, the coefficient α in equation (2) is fine tuned, for the buyer and vendor stages, to minimization overall cost (Chen 1999). Buyers and vendors incur costs in every period. Buyer costs include penalty cost, holding cost, and ordering cost. 2 Partial shipments can be an important factor, and are currently under study.
  7. 7. Vendor costs include holding cost and ordering cost. All costs are linear to the number of units of products. 4.1 Bullwhip Effect Bullwhip effect is been observed for all four models. Since there are multiple buyers and vendors in our models, orders of agents that belong to a same stage are aggregated into a stage order. Figure 2 shows that the degree of order variability increases up along the supply chain in model NR. The two graphs in the figure shows that bullwhip effect increases with physical lead-time. Bullwhip effect can be quantified by ratio of order variance of adjacent stages (Metters 1997). Since our models have three stages, two such ratios are computed. Ratio 1 is computed by dividing the variance of aggregated buyer stage orders by the variance of aggregated customer demands. And ratio 2 is computed by dividing variance of aggregated vendor stage orders by variance of aggregated buyer stage orders. To reduce the effect of initial conditions, data from the first 50 periods is discarded. Comparison between model NR and DS shows that marketization at supply chain level reduces bullwhip effect. Model NR’s ratio 1 (5.53) is significantly smaller (p < 0.0001) than that of model DS’ (6.47). Model NR’s ratio 2 (1.20) is marginally smaller (p =0.0606) than that of model DS’ (1.2). Comparison between model NR and DM shows that marketization at the industry level has no significant impact on bullwhip effect. Neither ratio 1 nor ratio 2 is significantly different between the two models. (See table 1 for all ratios) Figure 2: Bullwhip Effect A: Aggregated stage orders (model NR, physical delay [0 3], random seed =1):
  8. 8. Units 180 160 140 120 100 80 60 40 20 0 0 50 100 150 200 StepNo total_consumer_demand total_buyer_demand total_vendor_demand B: Aggregated stage orders (model NR, physical delay [10 20], random seed =1) Units 500 400 300 200 100 0 0 50 100 150 200 StepNo total_consumer_demand total_buyer_demand total_vendor_demand Together, the comparisons between model NR, DS and DM shows that it is the change from one vendor (as in model DS) to multiple vendors (as in model NR and DM), instead of the change from static customer-vendor relationships to dynamic (as in model NR) that reduces bullwhip effect. This, however, is contingent on the vendor selection mechanism. Comparison between model NI and DM shows that using inventory level as a vendor selection mechanism marginally (p=0.1008) increases bullwhip effect between the buyer stage and customers, but significantly (p<0.0001) reduces bullwhip effect between the vendor and buyer stages.
  9. 9. Table 1: Ratio of stage order variances Model Ratio 1 (stddev) Ratio 2 NR 5.53 (0.77) 1.20 (0.02) NI 5.65 (0.60) 1.15 (0.03) DS 6.47 (0.90) 1.22 (0.02) DM 5.37 (0.65) 1.21 (0.02) 4.2 Service Level Service level is measured by averaged (per period) buyer stage lost sales. Comparison between model NR, DS and DM shows that that marketization at both supply chain and industry level leads to lower service level (more lost sales, see table 2). Model NR’s lost sales are significantly more than model DS’s (p<0.0001) and model DM’s (p=0.0217). The vendor selection mechanism did not show a significant impact on service level. Model NI’s lost sales are also significantly higher than that of model DS (p<0.0001) and model DM (p=0.0219). Note that service level as measured by lost sales can be dependent on the no-partial shipment policy used for the experiments here. Table 2: Average lost sales per period Model Lost Sales (stddev) NR 54.4 (0.9) NI 54.6 (1.6) DS 51.9 (1.8) DM 53.8 (1.0) 4.3 Inventory Level Inventory level indicates efficiency of a supply chain. The more inventory a supply chain carries, the less efficient it is. It was found that marketization at both supply chain and industry level leads to lower vendor stage inventory levels (see figure 3a). Marketization was not found to be of impact on buyer stage inventory levels (see figure 3b). 5 Conclusion At individual agent level, the agent-based supply network tool developed in this study can incorporate information lead-time, production lead-time, and physical (transportation) lead-time. The tool can also model different production capacities, inventory-replenishment rules, and vendor selection rules for upstream agents. Different demand patterns can be modeled through the customer agent.
  10. 10. At the supply chain/industry level, the current tool can model a varying number of agents and stages and with different configurations between stages. With model configurations and parameters set in consultation with supply chain managers, face validation of models and emergent effects can be examined. When aligned with real world data, the tool allows supply chain managers to “experiment” with different business scenarios and observe results. Models constructed can be aligned at individual, stage, and supply chain/ industry levels. Figure 3a: Aggregated vendor stage inventory levels Units 40 30 20 10 0 10 20 30 RandomSeed DM DS NI NR Figure 3b: Aggregated buyer stage inventory levels Units 112 111 110 109 108 107 106 105 104 103 102 101 100 99 98 0 10 20 30 RandomSeed DM DS NI NR
  11. 11. Reference: Axtell, R. (2000). Why Agents? On the Varied Motivations for Agent Computating in the Social Sciences. The Brookings Institute, Center on Social and Economic Dynamics, Working Paper No. 17. Beamon, B. M. and V. C. P. Chen (2001). "Performance analysis of conjoined supply chains." International Journal of Production Research 39(14): 3195-3218. Chen, F. (1999). "Decentralized supply chains subject to information delays." Management Science 45(8): 1076-1090. Croom, S. (2001). "Restructuring supply chains through information channel innovation." International Journal of Operations & Production Management 21(4): 504-550. Forrester, J. (1961). Industrial Dynamics. New York, MIT Press. Fransoo, J. C. and M. J. F. Wouters (2000). "Measuring the bullwhip effect in the supply chain." Supply Chain Management 5(2): 78. Lee, H., V. Padmanabhan, and S. Whang (1997). "Information Distortion in a Supply Chain: The Bullwhip Effect." Management Science 43(4): 546-558. Malone, T., J. Yates, and R. Benjamin (1987). "Electronic markets and electronic hierarchies: effects of information technology on market structure and corporate strategies." Communications of the ACM 30(6): 484-497. Metters, R. (1997). "Quantifying the bullwhip effect in supply chains." Journal of Operations Management 15: 89-100. Parunak, H. V., R. Savit, and R. Riolo (1998). "Agent-Based Modeling vs. Equation- Based Modeling: A Case Study and Users’ Guide." Proceedings of Multi-agent systems and Agent-based Simulation (MABS'98): 10-25. Williamson, O. E. (1975). Markets and Hierarchies: Analysis and Antitrust Implications. New York, The Free Press.

×