Large inventory-service optimization in configure-to-order systems

Abstract

A manufacturing process is migrated from an existing operation to a configure-to-order (CTO) system. As the CTO operation will eliminate the “machine-type model” (MTM) inventory of the existing operation, the emphasis is shifted to the components, or “building blocks”, which will still follow the build-to-stock scheme, due to their long leadtimes, and hence still require inventory. The solution involves an inventory-service trade-off of the new CTO system, resulting in performance gains, in terms of reduced inventory cost and increased service level. Other benefits include better forecast accuracy through parts commonality and risk-pooling, and increased customer demand, as orders will no longer be confined within a restricted set of pre-configured MTMs.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention generally relates to computer assisted manufacturing processes and, more particularly, to a model for re-engineering a build-to-stock operation to a configure-to-order operation centered around “building blocks”, thereby keeping inventory only at the component level.

2. Background Description

A configure-to-order (CTO) system is a hybrid of make-to-stock and make-to-order operations: a set of components (subassemblies) are built to stock whereas the end products are assembled to order. This hybrid model is most suitable in an environment where the time it takes to assemble the end product is negligible, while the production/replenishment leadtime for each component is much more substantial. Personal Computer (PC) manufacturing is a good example of such an environment. By keeping inventory at the component level, customer orders can be filled quickly. On the other hand, postponing the final assembly until order arrival provides a high level of flexibility in terms of product variety, and also achieves resource pooling in terms of maximizing the usage of component inventory. Therefore, the CTO system appears to be an ideal business process model that provides both mass customization and a quick response time to order fulfillment.

Such a hybrid model is often referred to as an assemble-to-order (ATO) system in the research literature. In an ATO system, usually there is a pre-fixed set of end-product types from which customers must choose. In contrast, a CTO system takes the ATO concept one step further, in allowing each customer to configure his/her own product in terms of selecting a personalized set of components that go into the product. Aside from checking that the product so configured must “make sense”, there is no “menu” of product types that limits the customer's choice.

PC manufacturing traditionally has been a build-to-plan (or build-to-forecast) process, a process that is sometimes referred to as the “machine-type model” (MTM) operation. There is a set of end products, or MTMs. Demand forecasts over a future planning horizon are generated for each MTM, and updated periodically for each planning cycle, typically, a weekly cycle. A “materials requirements planning” (MRP) type explosion technique is then used to determine the requirements for the components over the planning horizon, based on the bill-of-materials (BOM) structure of each end product. Because of the random variation involved in demand forecasts, safety stock is usually kept for each end product, as well as at each component level, in order to meet a desirable customer service level. However, holding finished goods inventory for any length of time is very costly in the PC business, where product life cycle is measured in months and price reduction takes place almost every other week.

Y. Wang, “Service Levels in Production-Inventory Networks: Bottlenecks, Tradeoffs, and Optimization”, Ph.D. Dissertation, Columbia University, 1988, applies an asymptotic result in an optimization problem to minimize average inventory holding cost with a constraint on the order fill-rate. J. M. Swaminathan and S. R. Tayur, “Stochastic Programming Models for Managing Product Variety”, in Quantitative Models for Supply Chain Management, S. Tayur, R. Ganeshan and M. Magazine (eds.), Kluwer Academic Publishers, Norwell, 1999, pp. 585-624, use stochastic programming models to study three different strategies at the assembly stage; utilizing component commonality, postponement (the “vanilla box approach”), and integrating assembly task design and operations. Other related recent works, not necessarily in the CTO setting, include Y. Aviv and A. Federgruen, “The Benefits of Design for Postponement”, in Quantitative Models for Supply Chain Management, S. Tayur, R. Ganeshan and M. Magazine (eds.), Kluwer Academic Publishers, Norwell, 1999, pp. 553-584, A. Garg and H. L. Lee, “Managing Product Variety: An Operations Perspective”, in Quantitative Models for Supply Chain Management, S. Tayur, R. Ganeshan and M. Magazine (eds.), Kluwer Academic Publishers, Norwell, 1999, 467-490, L. Li, “The Role of Inventory in Delivery-Time Competition”, Management Science, 38 (1992), 182-197, and S. Mahajan and G. J. van Ryzin, “Retail Inventories and Consumer Choice”, in Quantitative Models for Supply Chain Management, S. Tayur, R. Ganeshan and M. Magazine (eds.), Kluwer Academic Publishers, Norwell, 1999, 491-552.

SUMMARY OF THE INVENTION

The problem solved by the present invention is to migrate a manufacturing process from an existing operation to a Web-based CTO operation where customer orders will be taken from the Internet. As the CTO operation will eliminate the MTM inventory, the emphasis will be shifted to the components, or “building blocks”, which will still follow the build-to-plan scheme, due to their long leadtimes, and hence still require inventory. The solution involves an inventory-service trade-off of the new CTO system, with resulting performance gains, in terms of reduced inventory cost and increased service level. There are other benefits: There will be better forecast accuracy through parts commonality and risk-pooling. Customer demand is expected to increase, as orders will no longer be confined within a restricted set of pre-configured MTMs.

Whereas most studies in the literature focus on certain segments of the supply chain, modeled as simple stand-alone queues, the present invention aims at modeling large-scale, end-to-end enterprise supply chains, such as those in the PC industry. The centerpiece of the invention is a network of inventory queues, which integrates inventory control with the delay-capacity features of queues. The network configuration is determined by the bill of materials (BOM) structure of the end products. The solution approach is a decomposition-based approximation, coupled with nonlinear optimization using conjugate gradient search. The model can be used not only to study the inventory-service tradeoff, but also to compare the effects on inventory and service performances through changing the network configuration; i.e., the supply chain structure, and to identify the best configuration.

BRIEF DESCRIPTION OF THE DRAWINGS

The foregoing and other objects, aspects and advantages will be better understood from the following detailed description of a preferred embodiment of the invention with reference to the drawings, in which:

FIG. 1 is a diagram illustrating an example of a bill-of-material structure of seven components used to produce three end products in an configure-to-order system;

FIGS. 2A and 2B, taken together, are a flow diagram of an implementation of the inventory optimization algorithm according to the invention;

FIGS. 3A and 3B, taken together, are a block diagram showing a hardware platform on which the invention may be implemented;

FIG. 4 is a graph of optimal days-of-supply for selected components in non-stationary demand environment for a target service level of 95%;

FIG. 5 is a graph showing a comparison between a build-to-stock (“as-is”) and an configure-to-order (“to-be”) system;

FIG. 6 is a graph showing the effect of improving forecast accuracy; and

FIG. 7 is a graph showing the effect of product variety on inventory.

DETAILED DESCRIPTION OF A PREFERRED EMBODIMENT OF THE INVENTION

We consider a hybrid model, by which each end product is configured to order from a set of components, which, in turn, are built to stock. In other words, no finished goods inventory is kept for any end product, whereas each component (“building block”) has its own inventory, replenished from a supplier following a base-stock policy.

Referring now to the drawings, and more particularly to FIG. 1, there is shown a diagram illustrating an example of this hybrid model. Abstractly, there are seven components comprising the set S={1, 2, . . . , 7} which are used in various combinations to produce three end products comprising the set M={1, 2, 3}. The components are maintained in inventory, but no inventory is maintained of end products. Each component inventory is indexed by i, iεS, where S denotes the set of all components. Associated with each component is a “store”, where the inventory is kept.

ATO Environment

First, consider the traditional ATO environment. That is, there is a pre-specified set of end products (i.e., machine type models), with each type indexed by mεM where M denotes the set of all product types. For each product type, there is a corresponding demand stream.

Time is discrete, indexed by t, with each time unit called a period. Let D_m(t) denote the demand of type m in period t. Each unit of type m demand requires a subset of components, denoted S_m⊂S, possibly with multiple units of each component. From the BOM structure of the end product, we can identify M_i as the subset of end products that all require component T as part of the assembly. That is, M_i={m: iεS_m}.

There are two kinds of leadtimes; those associated with the components, and those associated with the end-products:

• • L_iⁱⁿ, iεS: the in-bound leadtime—the time for the supplier of component i to replenish to store i once an order is placed. Assume this leadtime is known through a given distribution. For instance, a normal distribution with mean and variance given.
  • L_i^out, mεM: the out-bound leadtime—the time to supply a customer demand of type m, provided there is no stockout of any component iεS_m. This time includes the order processing time, the assembly/reconfiguration time, and the transportation time to deliver the order. The distribution of this leadtime is also assumed known.

We can now express the demand at store i in period t, for any iεS and any t=1, 2, . . . , as follows:

$D_i\left(t\right)=\sum _{m\in M_i}{D}_m\left(t+L_m^{\mathrm{out}}\right).$ Note this is the standard MRP type of demand-leadtime offset. We derive the mean and variance for D_i(t) as follows:

$\begin{array}{cc}E\left[D_i\left(t\right)\right]=\sum _{m\in M_i}\sum _{l}E\left[D_m\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right],& \left(1\right)\\ \mathrm{Var}\left[D_i\left(t\right)\right]=\sum _{m\in M_i}\sum _{l}E\left[D_m^2\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right]-\sum _{m\in M_i}{\left(\sum _{l}E\left[D_m\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right]\right)}^2& \left(2\right)\end{array}$ The variance calculation above assumes independence among demand classes; for correlated demands, simply add to the above the covariance terms.

CTO Environment

The above analysis is now adapted to the CTO environment. In this environment, there is no pre-specified product menu; in principle, every order can require a distinct set of components. Hence, we use M to denote the set of product/demand families that use the same set of components S. For instance, M={low-end machines, high-end machines, servers}, or M={individuals, small business, corporations}.

Suppose D_m is the demand associated with product family m in a certain period (for simplicity, the time argument t is omitted). Suppose each unit (order) in D_m requires a random number of units from component i, denoted as X_mi, which takes on non-negative integer values. The sets M and M_i are still defined as before, but they now take on the following form:S_m=S−{i:X_mi=0}, and M_i=M−{m:X_mi=0}

We can then translate the end-product demand into demand for each component i (per period).

$D_i=\sum _{m\in M_i}\sum _{k=1}^{D_m}{X}_{\mathrm{mi}}\left(k\right).$ where X_mi(k), for k=1, 2, . . . , are independent, identically distributed (i.i.d.) copies of X_mi. Assume the mean and the variance of X_mi are known. For instance, starting from the joint distribution of (X_mi)_iεSm, derived from empirical demand data, we can derive the marginal distributions, and then the mean and the variance of X_mi. Applying Wald's identity (S. M. Ross, Stochastic Processes, John Wiley & Sons, New York, 1996), we can derive:

$\begin{array}{cc}E\left[D_i\right]=\sum _{m\in M_i}E\left[X_{\mathrm{mi}}\right]E\left[D_m\right],& \left(3\right)\\ \begin{array}{c}\mathrm{Var}\left[D_i\right]=\sum _{m\in M_i}\left(E\left[D_m\right]\mathrm{Var}\left[X_{\mathrm{mi}}\right]+\mathrm{Var}\left[D_m\right]E^2\left[X_{\mathrm{mi}}\right]\right)\\ =\sum _{m\in M_i}(E^2\left[X_{\mathrm{mi}}\right]E\left[D_m^2\right]+\mathrm{Var}\left[X_{\mathrm{mi}}\right]E\left[D_m\right]-\\ E^2\left[X_{\mathrm{mi}}\right]E^2\left[D_m\right])\end{array}& \left(4\right)\end{array}$

Therefore, equations (1) and (2) can be generalized as follows:

$\begin{array}{cc}E\left[D_i\left(t\right)\right]=\sum _{m\in M_i}E\left[X_{\mathrm{mi}}\right]\sum _{l}E\left[D_m\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right],& \left(5\right)\\ \begin{array}{c}\mathrm{Var}\left[D_i\left(t\right)\right]=\sum _{m\in M_i}{E}^2\left(X_{\mathrm{mi}}\right)\sum _{l}E\left[D_m^2\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right]+\\ \sum _{m\in M_i}\mathrm{Var}\left(X_{\mathrm{mi}}\right)\sum _{l}E\left[D_m\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right]-\\ \sum _{m\in M_i}{E}^2\left(X_{\mathrm{mi}}\right){\left(\sum _{l}E\left[D_m\left(t+l\right)\right]P\left[L_m^{\mathrm{out}}=l\right]\right)}^2.\end{array}& \left(6\right)\end{array}$

Base-Stock Control

As mentioned above, suppose each store i follows a base-stock policy. Let R_i(t) denote the reorder point (or, base-stock level) in period t. Express the reorder point as follows:R_i(t):=μ_i(t)+k_i(t)σ_i(t),  (7)where k_i(t) is the desired safety factor, while μ_i(t) and σ_i(t) can be derived (via queuing analysis as follows:

$\begin{array}{cc}{\mu }_i\left(t\right)=\sum _{s=t}^{t+l_i^{\mathrm{in}}-1}E\left[D_i\left(s\right)\right],& \left(8\right)\\ {\sigma }_i^2\left(t\right)=\sum _{s=t}^{t+l_i^{\mathrm{in}}-1}\mathrm{Var}\left[D_i\left(s\right)\right],& \left(9\right)\end{array}$ where l_iⁱⁿ:=E[L_iⁱⁿ] is the expected in-bound leadtime, and E[D_i(s)] and Var[D_i(s)] follow equations (1) and (2), respectively. Note that since time is discrete, we shall round up any real-valued l_iⁱⁿ to the next integer to be used in equations (8) and (9). Also note that the reason we choose to write R_i(t) as in equation (7) is because we model the demand as follows:D_i(t)=μ_i(t)+σ_i(t)·Z,  (10)where Z denotes the standard normal variate.

To facilitate implementation, it is often desirable to translate R_i(t) into “days of supply” (DOS), or more precisely, periods of supply. To do so, note that the μ_i(t) part of R_i(t) simply translates into periods of demand. In addition, we can turn the k_i(t)σ_i(t) part of R_i(t) into

$\frac{k_i\left(t\right){\sigma }_i\left(t\right)}{\frac{{\mu }_i\left(t\right)}{l_i^{\mathrm{in}}}}$ periods of demand. Hence, we can express R_i(t) in terms of periods of demand, or DOS, as follows:

$\begin{array}{cc}{\mathrm{DOS}}_i\left(t\right)=l_i^{\mathrm{in}}\left[1+k_i\left(t\right)\frac{{\sigma }_i\left(t\right)}{{\mu }_i\left(t\right)}\right]& \left(11\right)\end{array}$ Note the intuitively appealing form of equation (11), in particular the safety-stock (or rather, safety time) part, which is equal to the product of the safety factor and the coefficient of variation (i.e., the ratio of standard deviation to mean) of the component demand over the (in-bound) leadtime.

For the inventory performance measures, the expected on-hand inventory and back-order level are as follows:E[I_i(t)]=σ_i(t)H(k_i(t)), E[B_i(t)]=σ_i(t)G(k_i(t)).  (12)Here the function G(•) is defined as:

$\begin{array}{cc}G\left(x\right):={E\left[Z-x\right]}^{+}={\int }_x^{\infty }\left(z-x\right)\varphi \left(z\right)dz=\varphi \left(x\right)-x\stackrel{\_}{\Phi }\left(x\right),& \left(13\right)\end{array}$ with Z denoting the standard normal variate, φ and Φ denoting, respectively, the density function and the distribution function of Z, and Φ:=1−Φ(x). The function H(•) is defined as:H(x):=E[x−Z]⁺=x+G(x)=φ(x)+xΦ(x).  (14)Note that the expressions in (12) are based on the following identities:I_i(t)=[R_i(t)−D_i(t)]⁺, and B_i(t)=[D_i(t)−R_i(t)]⁺;along with equations (7) and (10). Furthermore, we haveP[D_i(t)≧R_i(t)]=P[Z≧k_i(t)]= Φ(k_i(t)),  (15)which relates to the service requirement to be elaborated in the next section.

Next, suppose demand is stationary, i.e., for each demand class m, D_m(t) is invariant in distribution over time. Then, the mean and the variance of demand per period for each component i follow the expressions in (3) and (4). Consequently, (8) and (9) reduce to the followingμ_i=l_iⁱⁿE[D_i], and σ_i²=l_iⁱⁿVar[D_i].  (16)We can then writeR₁=l_iⁱⁿE[D_i]+k_i√{square root over (l_iⁱⁿ)}sd[D_i];  (17)and hence,

$\begin{array}{cc}{\mathrm{DOS}}_i=\frac{R_i}{E\left[D_i\right]}=l_i^{\mathrm{in}}+k_i{\theta }_i\sqrt{l_i^{\mathrm{in}}}=l_i^{\mathrm{in}}\left[1+k_i\frac{{\theta }_i}{\sqrt{l_i^{\mathrm{in}}}}\right],& \left(18\right)\end{array}$ where θ_i=sd[D_i]/E[D_i] is the coefficient of variation of the demand per period for component i. (Hence, θ_i/√{square root over (l_iⁱⁿ)} is the coefficient of variation of the demand over the leadtime l_iⁱⁿ, which is consistent with the general formula in equation (11).)

Sometimes it is more appropriate to adjust the demand distribution to account for non-negativity. Specifically, instead of D=μ+σZ, where Z is the standard normal variate, we should have {tilde over (D)}=[μ+σZ]⁺. The adjusted mean follows from equation (13) as follows:

$\begin{array}{cc}E\left[\tilde{D}\right]=\sigma {E\left[Z-\left(-\frac{\mu }{\sigma }\right)\right]}^{+}=\sigma G\left(-\frac{\mu }{\sigma }\right).& \left(19\right)\end{array}$ To derive the adjusted variance, note the following:

$\begin{array}{c}E\left\{{\left[{\left(Z-x\right)}^{+}\right]}^2\right\}={\int }_x^{\infty }{\left(z-x\right)}^2\varphi \left(z\right)dz\\ =x\varphi \left(x\right)+\stackrel{\_}{\Phi }\left(x\right)-2x\varphi \left(x\right)+x^2\stackrel{\_}{\Phi }\left(x\right)\\ =\stackrel{\_}{\Phi }\left(x\right)-\mathrm{xG}\left(x\right),\end{array}$ where the last equation makes use of equation (13). Hence,

$\begin{array}{cc}\begin{array}{c}\mathrm{Var}\left[\tilde{D}\right]={\sigma }^2\mathrm{Var}\left\{{\left[Z-\left(-\frac{\mu }{\sigma }\right)\right]}^{+}\right\}\\ ={\sigma }^{2}E\left\{{\left[{\left(Z-\left(-\frac{\mu }{\sigma }\right)\right)}^{+}\right]}^2\right\}-{\left[E\left(\tilde{D}\right)\right]}^2\\ ={\sigma }^2\left[\stackrel{\_}{\Phi }\left(-\frac{\mu }{\sigma }\right)+\frac{\mu }{\sigma }G\left(-\frac{\mu }{\sigma }\right)-G^2\left(-\frac{\mu }{\sigma }\right)\right].\end{array}& \left(20\right)\end{array}$ For moderately large x (say, x≧2), from equation (13), we have G(−x)≅x, and henceE[{tilde over (D)}]≅E[D], Var[{tilde over (D)}]≅Var[D], from equations (19) and (20). Therefore, the above adjustment is only needed when the coefficient of variation of the demand, σ/ν, is relatively large, say, 0.5 or above.

From equations (7), (17) and (18), it is clear that to identify the base-stock policy is tantamount to specifying the safety factor k_i for each component inventory. In the following sections, we discuss how to set the safety factor values so as to achieve the best inventory-service performance as specified in the optimization problems below.

For ease of exposition, we shall focus on stationary demand. For non-stationary demand, we can simply solve the optimization problems below period by period.

Service Requirement

To start with, consider the special case of each order of type m requires exactly one unit of component iεS_m. Let α be the required service level, defined here as the off-shelf availability of all the components required to configure a unit of type m product, for any m. Let E_i denote the event that component i is out of stock. Then, we require, for each end product mεM,P[∪_iεSmE_i]≦1−α.From the well-known inclusion-exclusion formula:

$P\left[{\bigcup }_{i\in S_m}{E}_i\right]=\sum _{i}P\left(E_i\right)-\sum _{i<j}P\left(E_i\bigcap E_j\right)+\sum _{i<j<k}P\left(E_i\bigcap E_j\bigcap E_k\right)-\dots ,$ we have, as an approximation,

$\begin{array}{cc}P\left[{\bigcup }_{i\in S_m}\right]\cong \sum _{i\in S_m}P\left(E_i\right)=\sum _{i\in S_m}\stackrel{\_}{\Phi }\left(k_i\right)\le 1-\alpha .& \left(21\right)\end{array}$

There is another way to arrive at the above inequality. Suppose we express the service requirement as follows:

$\begin{array}{cc}\prod _{i\in S_m}\Phi \left(k_i\right)\ge \alpha ,m\in M.& \left(22\right)\end{array}$ Note that the left hand side in inequality (22) is, in fact, a lower bound of the no-stockout probability of the set of components in S_m that is required to configure the end-product m; i.e., it is a lower bound of the desired off-shelf availability. This claim (of a lower bound) can be argued by using stochastic comparison techniques involving the notion of association. (Refer to, e.g., S. M. Ross, Stochastic Processes, 2^nd ed., Wiley, New York (1998), for background materials.) Intuitively, since the component inventories are driven by a common demand stream {D_m(t)}, and hence positively correlated, the chance of missing one or several components must be less than when the component inventories are independent, which is what is assumed by the product on the left hand side of inequality (22).

Since

$\prod _{i\in S_m}^{}\Phi \left(k_i\right)=\prod _{i\in S_m}^{}\left[1-\stackrel{\_}{\Phi }\left(k_i\right)\right]\cong 1-\sum _{i\in S_m}^{}\stackrel{\_}{\Phi }\left(k_i\right),$ combining the above and inequality (22), we arrive at the same inequality in (21).

In the general setting of CTO, consider demand of product family m. Let A⊂S_m denote a certain configuration, which occurs in this demand stream with probability P(A). Then the no-stockout probability, Π_iεAΦ(k_i), should be weighted by P(A). Hence, the service requirement in (22) should be changed to

$\begin{array}{cc}\begin{array}{c}\alpha \le \sum _{A\subseteq S_m}^{}P\left(A\right)\prod _{i\in A}^{}\Phi \left(k_i\right)\\ \approx \sum _{A\subseteq S_m}^{}P\left(A\right)\left[1-\sum _{i\in A}^{}\stackrel{\_}{\Phi }\left(k_i\right)\right]\\ =1-\sum _{A\subseteq S_m}^{}P\left(A\right)\sum _{i\in A}^{}\stackrel{\_}{\Phi }\left(k_i\right)\\ =1-\sum _{i\in S_m}^{}\left(\sum _{i\in A}^{}P\left(A\right)\right)\stackrel{\_}{\Phi }\left(k_i\right).\end{array}& \\ \mathrm{Since}& \\ \sum _{i\in A}^{}P\left(A\right)=P\left(X_{mi}>0\right):=r_{mi},& \left(23\right)\end{array}$ in the CTO environment the service requirement in (21) can be extended to the following:

$\begin{array}{cc}\sum _{i\in S_m}^{}{r}_{\mathrm{mi}}\stackrel{\_}{\Phi }\left(k_i\right)\le 1-\alpha ,& \left(24\right)\end{array}$ where r_mi follows (23).

Note that in the CTO environment, in particular when X_mi≧1, the stockout at component i should occur more often than Φ(k_i) due to the batch size associated with each order. In particular, (15) should be modified as follows (omitting the time argument t):

$P\left[D_i+X_{mi}\ge R_i\right]=P\left[Z\ge k_i-\frac{X_{\mathrm{mi}}}{{\sigma }_i}\right]=\stackrel{\_}{\Phi }\left(k_i-\frac{X_{\mathrm{mi}}}{{\sigma }_i}\right),$ which is larger than Φ(k_i). But this gap should be insignificant, since

$\frac{X_{\mathrm{mi}}}{{\sigma }_i}=\frac{X_{\mathrm{mi}}}{{\theta }_i{\mu }_i},$ where θ_i:=σ_i/μ_i is the coefficient of variation (or, forecast error); and the batch size of incoming orders is usually orders of magnitude smaller when compared against μ_i, which is the mean of demand summed over all product types mεM_i and over the leadtime. Also note that this underestimation of the stockout probability is compensated by the overestimation involved in (22), since the latter is a lower bound of the no-stockout probability.

We can now relate the above off-shelf availability requirement to the standard customer service requirements expressed in terms of leadtimes, W_m. Suppose the required service level of type m demand is:P[W_m≦w_m]α, mεM,  (25)where w_m's are given data.

Consider type m demand. We have the following two cases:

• • (i) When there is no stockout at any store iεS_m—denoting the associated probability as π_0m(t), the delay is simply L_i^out, the out-bound leadtime.
  • (ii) Suppose there is a stockout at one or several stores in the subset s⊂S_m. Denote the associated probability as πs_m(t). Then, the delay becomes L_i^out+τ_s, where τ_s is the additional delay before the missing components in s become available.Hence, we can write

$\begin{array}{cc}P\left[W_m\le w_m\right]={\pi }_{0m}\left(t\right)P\left[L_m^{\mathrm{out}}\le w_m\right]+\sum _{s\in S_m}^{}{\pi }_{\mathrm{sm}}\left(t\right)P\left[L_m^{\mathrm{out}}+{\tau }_s\le w_m\right].& \left(26\right)\end{array}$ In most applications, it is reasonable to assume thatL_m^out≦w_m and L_m^out+τ_s>w_m  (27)both hold almost surely. For instance, this is the case when the (nominal) outbound leadtime is nearly deterministic and shorter than what the customers require (this is, after all, a precondition for-running an configure-to-order operation); whereas the replenish leadtime for any component is substantially longer, hence it is virtually impossible to meet the customer service requirement in the case of any component stockout.

Clearly, with the assumed inequalities in (27), the first probability on the right hand side of (26) becomes one, and the probabilities under the summation all become zero. Consequently, the leadtime service requirement in (25) reduces to the off-shelf availability π_0m, and a lower bound of the latter is the left hand side of (22), as we explained earlier.

Inventory-Service Optimization

Our objective is to minimize the maximum stockout probability among all end-products subject to a given inventory budget. In view of equation (21) and the more general service contraint in (24), the problem can be presented as follows:

$\min \underset{m\in M}{\max }\left\{\sum _{i\in S_m}^{}{r}_{mi}\stackrel{\_}{\Phi }\left(k_i\right)\right\}$ $s.t.\sum _{i\in S}^{}{c}_i{\sigma }_{i}H\left(k_i\right)\le B,$ where c_i is the unit cost of the on-hand inventory of component, i, and B>0 is a given parameter, the budget limit on the expected overall inventory cost. Recall that σ_iH(k_i) is the expected on-hand inventory of component i (refer to equation (2)) and σ_i follows the specification in equation (16).

To solve the above optimization problem, we first apply a transformation of variables. Letx_i:= Φ(k_i), or k_i= Φ⁻¹(x_i)  (28)

Next, denoteg(x):=H( Φ⁻¹(x)).We can then reformulate the problem as follows:

$\begin{array}{cc}\min \beta & \left(29\right)\\ s.t.\sum _{i\in S_m}^{}{r}_{mi}{x}_i\le \beta ,m\in M& \left(30\right)\\ \sum _{i\in S}^{}{c}_i{\sigma }_{i}g\left(x_i\right)\le B& \left(31\right)\end{array}$ Here, the objective value β represents the upper limit on the stockout probability (or, “no-fill rate”) over any product types.

Several remarks are in order:

• • (i) Note that H(•) is an increasing and convex function, as evident from the first equality in equation (14), while is decreasing and convex (convex when the variable takes values on the non-negative half-line). Hence, g(•) is a decreasing and convex function.
  • (ii) For two-end-products, m and m′, if S_m′⊂S_m and r_m·i≦r_mi for iεS_m′, then the constraint corresponding to m′ becomes superfluous. We assume in the above formulation, all such superfluous constraints have already been removed through preprocessing: (Note, however, the demand for such end-products (as type m′) will still contribute to the calculation of σ_i in inequality (31), if m′εM_i.)
  • (iii) The constraint in inequality (31) is always binding, since g(•) is decreasing as explained in (i).

The Lagrangian corresponding to the above optimization problem is:

$L=B+\sum _{m\in M}^{}{\lambda }_m\left(\sum _{i\in S_m}^{}{r}_{mi}{x}_i-\beta \right)+\eta \left[\sum _{i\in S}^{}{c}_i{\sigma }_{i}g\left(x_i\right)-B\right],$ where λ_m>0 (mεM) and η≧0 are the Lagrangian multipliers. Hence, the optimal solution can be obtained from solving the following system of non-linear equations:

$\begin{array}{cc}\sum _{m\in M}^{}{r}_{mi}{\lambda }_m=-\eta c_i{\sigma }_i{g}^{\prime }\left(x_i\right),i\in S& \left(32\right)\\ \sum _{m\in M}^{}{\lambda }_m=1& \left(33\right)\\ \sum _{i\in S}^{}{c}_i{\sigma }_{i}g\left(x_i\right)=B& \left(34\right)\\ \sum _{i\in S_m}^{}{r}_{mi}{x}_i=\beta ,m\in M\mathrm{and}{\lambda }_m>0& \left(35\right)\end{array}$

As there is no easy way to solve the above system of non-linear equations, we propose a greedy algorithm, which works as follows: We first discretize the problem—let Δ be the step size. For instance, set Δ=0.01% (one basis points), and start with x_i=0 for all i. At each step, we identify the variable x_i*, such that increasing x_i* to x_i*+Δ will yield the smallest increase in the objective value. This corresponds to identifying whether the maximum, over all mεM, of the left hand side of inequality (30) will either (i) remain the same or (ii) increase by a positive amount.

Consider case (i) first, i.e., there is a subset, denoted A⊂S, such that for any iεA, we can increase x_i to x_i+Δ, without affecting the current β value in the algorithm. In this case, we pick the index i*εA as follows:

$\begin{array}{cc}{i}^{*}=\arg \underset{i\in A}{\max }\left\{c_i{\sigma }_i\left[g\left(x_i\right)-g\left(x_i+\Delta \right)\right]/r_{mi}\right\}.& \left(36\right)\end{array}$ In the complement case (ii), increasing x_i (by Δ) for any iεS will result in a positive increase in the β value. In this case, we still follow (36) to pick the index i*, replacing A by S. That is, we choose i* such that it will result in the smallest increase in β but in the largest reduction in inventory budget.

The above procedure is continued until the inventory budget constraint is satisfied; i.e., until the left hand side of inequality (31) is brought down to B or less.

The main computation involved in each step of the above algorithm is to evaluate the incrementc_iσ_i[g(x_i)−g(x_i+Δ)].Since Δ is a small increment, we can approximate the above difference by the derivative,c_iσ_i[−g′(x_i+Δ/2)]·Δ.Note that from equation (10), we have

$\begin{array}{cc}{H}^{\prime }\left(x\right)=-x\varphi \left(x\right)+\Phi \left(x\right)+x\varphi \left(x\right)=\Phi \left(x\right).\mathrm{And},{\left[{\stackrel{\_}{\Phi }}^{-1}\left(x\right)\right]}^{\prime }=\frac{-1}{\varphi \left({\stackrel{\_}{\Phi }}^{-1}\left(x\right)\right)}.\mathrm{Hence},-g^{\prime }\left(x\right)=-\Phi \left({\stackrel{\_}{\Phi }}^{-1}\left(x\right)\right)·\frac{-1}{\varphi \left({\stackrel{\_}{\Phi }}^{-1}\left(x\right)\right)}=\frac{1-x}{\varphi \left({\stackrel{\_}{\Phi }}^{-1}\left(x\right)\right)}.& \left(37\right)\end{array}$

To summarize, the algorithm is implemented as illustrated in the flow diagram of FIG. 2. The process begins with an initialization phase in function block 201 in which:

For each iεS, set x_i:=0.

For each iεS, set r_mi:=P(X_mi>0)

For each mεM, set β_m:=0.

Set β:=0.

At this point, the process enters the outer one of three nested processing loops. The first step is to initialize the set of active building blocks, A, to A={ } in function block 202. Then the second processing loop is entered in function block 203 which considers each iεS, followed by entering the third processing loop in function block 204 which considers each end product m that uses component i in its bill-of-material. In function block 205, set β_m:=β_m+r_miΔ, for all m such that iεS_m to perform the processing in the third nested processing loop. Then, in function block 206, compute δ_i:=max_m{β_m}−β. A determination is made in decision block 207 to determine if δ_i≦0. If so, then add i to the set of active building blocks, A:=A+{i}, to complete the processing in the second nested processing loop. In function block 208, determine if the set of active components A is non-empty; if so, set B:=A. otherwise set B:=S where B is a set of component indexes. In function block 209, find i*:=arg max_iεB{−c_iσ_i/r_mig′(x_i+Δ/2)}, where −g′ (•) follows equation (37). In function block 210, set x_i*:=x_i*+Δ to update the no-stockout probability of component i*. In function block 211, compute β=max_mεMβ_m, and in function block 212, check whether inequality (31) is satisfied. If yes, stop; otherwise, in function block 213 update β_m for each mεM_i*, set β_m:=β_m+r_miΔ, and go to function block 202 to complete the first processing loop.

When there is only a single product type, the above algorithm is guaranteed to yield the optimal solution (modulo the discretization). For multiple product types, to solve the problem to optimality would require, at each step, the examination of all subsets of S, instead of just the individual components iεS. For instance, if components i and j are involved in two different product types, then it is possible, at some step, to increase both x_i and x_j to x_i+Δ and x_j+Δ without increasing the objective value.

Another case of interest is when each end product has (at least) one unique component, which we denote as im for product m. That is, M_im={i_m} is a singleton set. In this case, it is quite easy to check the solution returned by the greedy algorithm to determine whether the solution satisfies the optimality conditions in equations (32) to (35). To do so, first note that equation (34) is always satisfied. Furthermore, equation (35) is also automatically satisfied in this case. For if Σ_iεSir_mix_i<β for some end product lεM, then we can always increase the value of x_i, that corresponds to the unique component i_l; and this will only decrease the overall inventory budget, since g(x_it) is decreasing in x_it. Hence, we only need to check equations (32) and (33). From equation (32), taking into account M_il={i_m}, we obtain

${\lambda }_m=\eta \frac{c_{i_m}{\sigma }_{i_m}{g}^{\prime }\left(x_{i_m}\right)}{r_{{\mathrm{mi}}_m}},m\in M,$ which, along with equation (33), yields

$\eta ={\left[\sum _{m\in M}^{}\frac{c_{i_m}{\sigma }_{i_m}{g}^{\prime }\left(x_{i_m}\right)}{r_{{mi}_m}}\right]}^{-1}.$ What then remains is to check whether or not the λ_m's and η derived from the above, along with the x_i's returned from the algorithm, satisfy equation (32), for iεS\{i_m, mεM}, the non-unique components.

The “dual” to the above optimization problem is to minimize the expected total inventory cost, subject to the stockout probability of each product no greater than β, a given parameter. That is,

$\begin{array}{cc}\min \sum _{i\in S}^{}{c}_i{\sigma }_{i}H\left({\stackrel{\_}{\Phi }}^{-1}\left(x_i\right)\right)& \left(38\right)\\ s.t.\sum _{i\in S_m}^{}{r}_{\mathrm{mi}}{x}_i\le \beta ,m\in M& \left(39\right)\end{array}$ The greedy heuristic described earlier still applies. The only difference is the stopping criterion. Here, we stop when some (if not all) of the constraints in (39) become binding such that no x_i values can be further increased. (That is, the set of end products corresponds to these binding constraints spans over the entire set of components S.)

FIG. 3 illustrates in block diagram form a hardware platform on which the algorithm illustrated in the flow diagram of FIG. 2 may be implemented. There are illustrated two databases 301 and 302 which provide inputs to processor 303. Database 301 contains optimization control parameters; specifically, budget limit on expected overall inventory cost and step size for the greedy algorithm. Database 302 contains manufacturing system parameters. These are (i) demand for end product mεM in period t=1, 2, . . . , T, (ii) inbound leadtime distribution for component iεS, (iii) outbound leadtime distribution for end product mεM, (iv) the distribution of the number of units of component i used in the assembly of end product mεM, (v) unit cost of component iεS, and (5) bill-of-materials for end product mεM. The processor 303 accesses these databases and first performs preprocessing steps of propagating demand using equations (5) and (6) and computing the mean and variance of the demand over the leadtime using equation (16). After the preprocessing, the processor performs optimization processing. This includes applying variable transformation using equation (28), reformulating the optimization problem using equations (29) to (31), and solving the optimization problem through the greedy algorithm as illustrated in the flow diagram of FIG. 2. Finally, the processor enters the postprocessing phase where the reorder points are computed using equations (5), (6), (17), and (28) and the reorder points are translated into days-of-supply using equation (18). The processor 303 outputs the computed information to an output device or database 304, this information including output days-of-supply for each component iεS and output achieved service level for each end product mεM.

Numerical Results

The purpose of this section is to illustrate numerically the effectiveness of the model and the algorithm developed above.

Problem Data

Consider a family of desktop computers that consists of six different end products, denoted PC 1-PC 6, which are configured from a set of 17 different components, or building blocks. The components are obtained from external suppliers with different supply leadtimes. Table 1 summarizes the data used in the numerical study, including the unit cost of each component and the supply leadtime (including time for shipment).

Each end product is configured from a set of building blocks as defined by its bill-of-material (BOM). The BOM structure is displayed in Table 2. The final assembly process, for each end product, takes no more than 1 to 2 hours, which is order-of-magnitude shorter than the component leadtime. So this fits well with the scenario discussed above.

We remark that the data provided in Tables 1 and 2 are based, on real data from the PC assembly system of a major computer manufacturer. To protect company confidentiality, we have altered the values of the data. The modified values, however, are still relatively close to a realistic product representation.

We examine the performance of our model under several parameter settings. The parameters varied include the budget limit on the expected inventory cost, demand variability (in terms of the coefficient of variation), and the BOM structure. We also investigate the performance of the model when applied in a non-stationary demand environment. We check the results obtained from the optimization against Monte-Carlo simulation. Simulation point estimates and confidence intervals were constructed from twenty independent replications with 1,000 time periods in each replication.

TABLE 1
Unit Costs and Supply Leadtimes of the Components
	Cost		Cost	Leadtime
Product	(in $)	Component	(in $)	(in days)
PC 1	1099	Memory 64 MB	290	15
PC 2	1331	Processor 450 MHZ	307	12
PC 3	1308	Processor 500 MHZ	538	12
PC 4	1243	Processor 550 MHZ	391	12
PC 5	1243	Processor 600 MHZ	799	12
PC 6	1098	Shell	42	5
		Shell (Common Parts 1)	114	8
		Shell (Common Parts 2)	114	8
		Hard Drive 6.8 GB	155	18
		Hard Drive 13.5 GB	198	18
		Hard Drive (Common Parts)	114	8
		CD ROM	43	10
		CD ROM (Common Parts)	114	10
		Video Graphics Card	114	6
		Ethernet Card	114	10
		Software Pre-Load 1	114	4
		Software Pre-Load 2	114	4

TABLE 2
BOM Structure
	Component Usage
	PC 1	PC 2	PC 3	PC 4	PC 5	PC 6
Memory 64 MB	1	1	1	1	1	1
Processor 450 MHZ	1
Processor 500 MHZ		1
Processor 550 MHZ			1	1	1
Processor 600 MHZ						1
Shell	1	1	1	1	1	1
Shell (Common Parts 1)	1	1	1	1	1	1
Shell (Common Parts 2)	1	1	1	1	1	1
Hard Drive 6.8 GB	1	1				1
Hard Drive 13.5 GB			1	1	1
Hard Drive (Common Parts)	1	1	1	1	1	1
CD ROM			1
CD ROM (Common Parts)			1
Video Graphics Card	1	1	1	1	1
Ethernet Card			1
Software Pre-Load 1	1	1	1		1
Software Pre-Load 2				1

Demand Variability

We first study the effect of demand variability on the service level, inventory budget, and optimal base-stock levels. Assume that demand for the six end products is independently identically distributed among the products and over the time, with a mean of E(D_m)=100 per period; the coefficient of variation is 0.2, 0.4, and 1.0. The budget limits are varied so as to obtain service levels (no-stockout probabilities) at about 90%, 95%, and 98%. The basic time period in all cases is one day.

Tables 3 to 5 summarize the numerical results for these cases. The column labeled “opt.” shows the optimal no-stockout probabilities of the end products for t e given inventory budget. The column labeled “sim.” shows the corresponding 95% confidence interval obtained from simulation. Input to the simulation are the base-stock levels for the building blocks generated by the optimization model. The base-stock level of each end product is zero, since each end product follows aft configure-to-order operation.

TABLE 3
Comparison for End-Product Service Levels;
E(Dm) 100; cv(Dm) = 0.2
	B = 800,000	B = 900,000	B = 1,000,000
	sim.	opt.	sim.	opt.	sim.	opt.
PC 1	0.949 ± 0.010	0.923	0.975 ± 0.006	0.963	0.989 ± 0.004	0.984
PC 2	0.949 ± 0.007	0.923	0.975 ± 0.005	0.963	0.989 ± 0.004	0.984
PC 3	0.980 ± 0.004	0.923	0.990 ± 0.003	0.963	0.995 ± 0.002	0.984
PC 4	0.966 ± 0.005	0.923	0.982 ± 0.004	0.963	0.991 ± 0.003	0.984
PC 5	0.991 ± 0.003	0.945	0.995 ± 0.003	0.974	0.997 ± 0.002	0.988
PC 6	0.945 ± 0.007	0.923	0.973 ± 0.005	0.963	0.988 ± 0.004	0.983

TABLE 4
Comparison for End-Product Service Levels;
E(Dm) 100; cv(Dm) = 0.4
	B = 1,500,000	B = 1,750,000	B = 2,000,000
	sim.	opt.	sim.	opt.	sim.	opt.
PC 1	0.927 ± 0.013	0.892	0.969 ± 0.007	0.956	0.989 ± 0.004	0.984
PC 2	0.928 ± 0.008	0.892	0.970 ± 0.006	0.956	0.989 ± 0.004	0.984
PC 3	0.969 ± 0.006	0.892	0.987 ± 0.003	0.956	0.995 ± 0.003	0.984
PC 4	0.949 ± 0.007	0.892	0.978 ± 0.004	0.956	0.990 ± 0.003	0.984
PC 5	0.983 ± 0.004	0.923	0.993 ± 0.003	0.969	0.997 ± 0.002	0.988
PC 6	0.919 ± 0.009	0.892	0.967 ± 0.005	0.955	0.988 ± 0.004	0.983

TABLE 5
Comparison for End-Product Service Levels;
E(Dm) 100; cv(Dm) = 1.0
	B = 800,000	B = 900,000	B = 1,000,000
	sim.	opt.	sim.	opt.	sim.	opt.
PC 1	0.926 ± 0.013	0.923	0.951 ± 0.011	0.957	0.967 ± 0.009	0.977
PC 2	0.926 ± 0.013	0.923	0.949 ± 0.008	0.957	0.967 ± 0.007	0.977
PC 3	0.954 ± 0.007	0.923	0.970 ± 0.005	0.957	0.981 ± 0.004	0.977
PC 4	0.933 ± 0.008	0.923	0.953 ± 0.007	0.957	0.967 ± 0.006	0.977
PC 5	0.971 ± 0.006	0.945	0.981 ± 0.005	0.970	0.988 ± 0.004	0.984
PC 6	0.928 ± 0.011	0.923	0.953 ± 0.009	0.957	0.969 ± 0.008	0.977

Observe that the no-stockout probabilities returned from the optimization model is generally conservative; i.e., lower than the simulated values. This is expected as the off-shelf availability of each end product used in the optimization model is a lower bound of the true value, as explained above. Also observe that the bound becomes tighter as the service level increases, along with the increased budget. Furthermore, the bound appears to be closer to the simulation results as the coefficient of variation (of demand) increases.

Among the six end-products, PC 5 is the only one that does not have a unique component. Observe that it alone achieves a higher service level than the other five end-products, which all have the same service level. In other words, all end-products with unique components have binding constraints in inequality (24), whereas PC 5 exceeds this service level.

For the same cases, Table 6 summarizes the comparison for the expected overall inventory costs. It is interesting to note that while the simulation runs result in higher service levels, the associated inventory costs are also slightly higher. (The relative error, however, is still well below 5%.) This suggests that in applications, when the inventory budget is a “hard” constraint, we need to input into the optimization model an inventory budget limit that is about 5% lower than its real value.

TABLE 6
Comparison for Expected Inventory Costs;
E(Dm) = 100
cv(Dm) = 0.2	cv(Dm) = 0.4	cv(Dm) = 1.0
opt.	sim.	rel. err.	opt.	sim.	rel. err.	opt.	sim.	rel. err.
800,000	841,819	4.9%	1,500,000	1,554,728	2.5%	4,000,000	4,056,800	1.4%
900,000	923,893	3.5%	1,750,000	1,774,716	1.4%	4,400,000	4,419,355	0.5%
1,000,000	1,014,657	1.4%	2,000,000	2,010,165	0.4%	4,800,000	4,792,957	0.1%

Table 7 displays the solution found by the optimization model for the case of cv(D_m)=1.0. For every component, the optimal solution is shown in three forms: base-stock level R_i* safety factor k_i*, and days-of-supply DOS_i. Clearly, these (optimal) values all increase as, the overall budget increases. The results also illustrate the effect of risk pooling: the higher the degree of commonality of a component, the lower the amount of safety stock required to maintain the service level. For instance, the safety stock for the 64 MB memory module, which is used in all six products, varies between 3.86 and 4.48 days-of-supply, whereas the safety stock for the ethernet card, which is unique to PC 3, is roughly twice that amount.

TABLE 7
Optimal Solution in Terms of Base-stock Levels, Safety Factors, and Days-of-Supply;
E(Dm) = 100; cv(Dm) = 1.0
	B = 4,000,000	B = 4,400,000	B = 4,800,000
	Ri*	ki*	DOSi	Ri*	ki*	DOSi	Ri*	ki*	DOSi
Memory 64 MB	11318	2.44	3.86	11509	2.64	4.18	11689	2.83	4.48
Processor 450 MHz	1760	1.62	5.60	1856	1.89	6.55	1950	2.16	7.49
Processor 500 MHz	1760	1.61	5.59	1856	1.89	6.55	1950	2.16	7.49
Processor 550 MHz	4793	1.99	3.98	4940	2.23	4.46	5088	2.48	4.96
Processor 600 MHz	1745	1.57	5.45	1842	1.85	6.42	1936	2.12	7.35
Shell	4803	3.29	3.00	4880	3.43	3.13	4940	3.54	3.23
Shell (Common Parts 1)	6818	2.91	3.36	6922	3.06	3.54	7045	3.24	3.74
Shell (Common Parts 2)	6818	2.91	3.36	6922	3.06	3.54	7045	3.24	3.74
Hard Drive 6.8 GB	7251	2.52	6.17	7390	2.71	6.63	7540	2.91	7.13
Hard Drive 13.5 GB	7028	2.21	5.42	7192	2.44	5.97	7356	2.66	6.52
Hard Drive (Common Parts)	6818	2.91	3.36	6942	3.09	3.57	7045	3.24	3.74
CD ROM	1863	2.73	8.62	1921	2.91	9.21	1978	3.09	9.77
CD ROM (Common Parts)	1742	2.35	7.42	1809	2.56	8.08	1877	2.77	8.76
Video Graphics Card	4538	2.81	3.08	4638	2.99	3.27	4729	3.16	3.46
Ethernet Card	1742	2.35	7.42	1809	2.56	8.0	1877	2.77	8.76
Software Pre-Load 1	2740	2.85	2.85	2805	3.01	3.01	2878	3.19	3.19
Software Pre-Load 2	795	1.97	3.94	842	2.21	4.41	889	2.44	4.88

Unique Components

Next, we illustrate the special case that each end-product involves at least one unique component, in particular, the verification of the optimality equations (26) to (29), as discussed at the end of the last section. Recall, in the above example, PC 5 is the only end-product that does not have a unique component. To obtain a scenario where each end product has at least one unique component, we modify the original bills-of-materials structure by letting PC 3 and PC 4 both use the 6.8 GB hard drive, while reserving the 13.5 GB hard drive as a unique component for PC 5. This way, PC 5, like all the other end-products, also has unique component.

We run the greedy algorithm on this modified BOM structure for the case of cv(D_m)=1.0. The results are summarized in Tables 8 and 9. Notice that the six end products now achieve the same service level because they all have unique components, and thus all constraints in inequality (24) become binding.

TABLE 8
Comparison for End-Product Service Levels
(With Each End-Product Having at Least One Unique Component);
E(Dm) = 100; cv(Dm) = 1.0
	B = 4,000,000	B = 4,400,000	B = 4,800,000
	sim.	opt.	sim.	opt.	sim.	opt.
PC 1	0.934 ± 0.012	0.933	0.956 ± 0.010	0.963	0.971 ± 0.008	0.981
PC 2	0.932 ± 0.009	0.933	0.956 ± 0.007	0.963	0.971 ± 0.006	0.981
PC 3	0.961 ± 0.006	0.933	0.975 ± 0.005	0.963	0.985 ± 0.003	0.981
PC 4	0.940 ± 0.007	0.933	0.958 ± 0.006	0.963	0.972 ± 0.005	0.981
PC 5	0.946 ± 0.009	0.933	0.965 ± 0.008	0.963	0.978 ± 0.007	0.981
PC 6	0.936 ± 0.010	0.933	0.959 ± 0.009	0.963	0.973 ± 0.007	0.981

When comparing the results for the original and modified BOM structures as shown in Tables 7 and 9, we find that the optimization produces similar days-of-supply targets for all components except the two hard drives. The safety stock of the 13.5 GB hard drive increases from 5.42 to 8.19 days-of-supply, whereas that of the 6.8 GB hard drive decreases from 6.17 to 5.05 days-of-supply. This again is a consequence of risk-pooling because modifying the BOM structure increases the commonality of the high-capacity hard drive, and at the same time decreases the commonality of the low-capacity hard drive.

To verify the optimality equations, recall from the discussion above that we first determine the Lagrangian multipliers, λ_m and η, based on the solutions returned by the algorithm, and then check the equations in (26) for all the non-unique components. Table 10 lists the gap, labeled “abs.err.”, between the left hand side and the right hand side of the equations in (26) for all the non-unique components. The gap is indeed very small, on the order of 10⁻⁸ or below. This clearly indicates that in this case the greedy algorithm has returned an optimal solution. (That there is a gap at all is mainly due to discretization.) The other two cases, with the coefficients of variation being 0.2 and 0.4, yield similar results.

TABLE 9
Optimal Solution in Terms of Base-Stock Levels, Safety Factors, and
Days-of-Supply (With Each End-Product Having at Least One Unique Component);
E(Dm) = 100; cv(Dm) = 1.0
	B = 4,000,000	B = 4,400,000	B = 4,800,000
	Ri*	ki*	DOS	Ri*	ki*	DOS	Ri*	ki*	DOS
Memory 64 MB	11357	2.48	3.93	11542	2.68	4.21	11731	2.88	4.55
Processor 450 MHz	1781	1.68	5.81	1876	1.95	6.75	1970	2.22	7.69
Processor 500 MHz	1781	1.68	5.81	1876	1.95	6.75	1970	2.22	7.69
Processor 550 MHz	4788	1.98	3.96	4936	2.23	4.45	5088	2.48	4.96
Processor 600 MHz	1767	1.63	5.66	1863	1.91	6.62	1956	2.18	7.56
Shell	4837	3.35	3.06	4940	3.54	3.23	5038	3.72	3.40
Shell (Common Parts 1)	6843	2.95	3.40	6963	3.12	3.60	7080	3.29	3.80
Shell (Common Parts 2)	6843	2.95	3.40	6963	3.12	3.60	7080	3.29	3.80
Hard Drive 6.8 GB	11525	2.66	5.05	11703	2.85	5.40	11881	3.04	5.76
Hard Drive 13.5 GB	2619	1.93	8.19	2724	2.18	9.23	2827	2.42	10.26
Hard Drive (Common Parts)	6843	2.95	3.40	6963	3.12	3.60	7080	3.29	3.80
CD ROM	1860	2.72	8.59	1921	2.91	9.21	1978	3.09	9.77
CD ROM (Common Parts)	1739	2.33	7.38	1807	2.55	8.06	1877	2.77	8.76
Video Graphics Card	4561	2.85	3.12	4663	3.04	3.33	4750	3.19	3.50
Ethernet Card	1739	2.33	7.38	1807	2.55	8.06	1877	2.77	8.76
Software Pre-Load 1	2752	2.88	2.88	2825	3.06	3.06	2896	3.24	3.24
Software Pre-Load 2	793	1.96	3.92	842	2.21	4.41	890	2.45	4.89

TABLE 10
Optimization Results for Non-Unique Components
(Scenario Where Each End Product Has at Least One Unique Component);
E(Dm) = 100; cv(Dm) = 1.0
	B = 4,000,000	B = 4,400,000	B = 4,800,000
	xi*	abs. err.	xi*	abs. err.	xi*	abs. err.
Memory 64 MB	0.0073	5.20−8	0.0041	3.04−8	0.4983	1.68−8
Processor 550 MHZ	0.0233	3.02−8	0.0129	1.84−8	0.4949	1.10−8
Shell	0.0005	5.59−8	0.0003	3.78−8	0.4999	2.26−8
Shell (Common Parts 1)	0.0018	5.31−8	0.0011	3.20−8	0.4996	1.73−8
Shell (Common Parts 2)	0.0018	5.31−8	0.0010	3.20−8	0.4996	1.73−8
Hard Drive 6.8 GB	0.0059	5.21−8	0.0034	3.10−8	0.4987	1.70−8
Hard Drive (Common Parts)	0.0018	5.31−8	0.0011	3.20−8	0.4996	1.73−8
Video Graphics Card	0.0025	4.08−8	0.0014	2.58−8	0.4994	1.24−8
Software Pre-Load 1	0.0022	2.72−8	0.0013	1.76−8	0.4995	0.93−8

Non-Stationary Demand

We next examine the effectiveness of the algorithm in handling non-stationary demand. For instance, when the demand forecast changes weekly, we can run the optimization model at the beginning of each week using the demand offset scheme described above. The base-stock levels are then updated weekly, based on the outcome of the optimization.

To illustrate this approach, consider the product structure described in Table 1. Consider a planning horizon of 100 weeks, over which the mean demand for each end-product changes weekly, while the coefficient of variation (which measures the forecast error) remains constant. Specifically, we randomly generate the weekly demand from a uniform sample over the interval [525,875], while maintaining the coefficient of variation at 2.646. Using the algorithm that solves the “dual” problem described in (31) and (32), we minimize the total expected inventory cost subject to a set of target no-stockout probabilities, 90%, 95%, and 98% (corresponding to β=0.10, 0.05, and 0.02). Table 11 compares the achieved service levels and the total expected inventory cost between simulation and the optimization model. Input to the simulation are the component base-stock levels, by week, generated by the optimization algorithm.

TABLE 11
Comparisons for End-product Service Levels and Total
Inventory Cost in Non-Stationary Demand Environment
	β = 0.10	β = 0.05	β = 0.02
	sim.	opt.	sim.	opt.	sim.	opt.
PC 1	0.928 ± 0.009	0.900	0.952 ± 0.007	0.950	0.971 ± 0.004	0.980
PC 2	0.926 ± 0.007	0.900	0.949 ± 0.005	0.950	0.969 ± 0.003	0.980
PC 3	0.944 ± 0.006	0.900	0.962 ± 0.004	0.950	0.975 ± 0.003	0.980
PC 4	0.929 ± 0.007	0.900	0.951 ± 0.006	0.950	0.968 ± 0.004	0.980
PC 5	0.950 ± 0.006	0.900	0.965 ± 0.004	0.950	0.978 ± 0.003	0.980
PC 6	0.922 ± 0.009	0.900	0.950 ± 0.008	0.950	0.970 ± 0.005	0.980
Total cost	$4,039,913	$3,830,505	$4,505,698	$4,329,152	$5,077,240	$4,924,748

For 90% and 95% service level targets, the simulated no-stockout probabilities do exceed the values returned by the optimization. At 98% service level, the simulated values are slightly below their analytical counterparts. This is largely due to the weekly update of the base-stock levels: whenever there is an increase in demand, and hence an increase in base-stock levels, from one week to the next, there is a time lag in building up the inventory, due to the supply leadtime, and the available stock may be insufficient to accommodate the increase in demand. This shortfall can be eliminated if the base-stock levels are updated more frequently, say, daily. We further observe that the analytical estimates of the expected total inventory cost are within 5.5% of the simulated values.

FIG. 4 shows the optimal safety-stock levels (i.e., base-stock minus work-in-progress (WIP)) converted to days-of-supply (DOS), for a subset of components in the example. Notice that the DOS values fluctuate very little over time. This suggests, in situations where the forecast error (in terms of its coefficient of variation) remains relatively constant over time, it suffices to follow a simple policy that tracks, for each component inventory, a constant DOS target (instead of updating the base-stock levels, which keep changing from period to period). Further, the results clearly demonstrate the benefit of risk-pooling: building blocks that go into a larger number of end products (e.g., memory modules) require much less safety stock than building blocks that are only used in a small number of end products (e.g., CD ROMs).

Process Re-Engineering Application

Here we describe the study mentioned earlier, which was part of a larger project aimed at the re-engineering of an existing business process from a build-to-stock operation to an assembly-to-order operation centered around “building blocks” (i.e., keeping inventory only at the component level). To carry out the study, we have developed two basic models: the “as-is” model that is a reflection of the present operation, and the “to-be” model that is based on the optimization model described in the previous sections—in particular, with the component inventory levels generated by the algorithm. For both models, we aggregate the business' production-inventory system into two stages, the first stage consists of the components, or building blocks, and the second stage includes the assembly and the order fulfillment.

Three factors have been identified as focal points of our study:

• • (i) manufacturing strategy—the “as-is” operation versus the “to-be” model;
  • (ii) the accuracy of demand forecast at the end-product level versus at the component level; and
  • (iii) the effect of mass customization as a result of, for instance, direct sales over the internet.

To study the first factor, we select a major product family at the business, which consists of 18 end-products configured from a total of 17 components. We use the business' existing data—including BOM, unit costs, assembly and procurement leadtimes—to run a detailed simulation model. The demand for each end-product is statistically generated, based on historical data. The days-of-supply targets are set to meet a required service level of 95% (for all end-products). Following the business' current practice, these targets are determined using a combination of simple heuristic rules and judgement calls from product managers, and verified by simulation (via trial and error).

We then feed the same data, including the same statitically generated demand streams, into the optimization model, which eliminates the finished-goods inventory at the end-product level and optimally sets the base-stock level for each component inventory. The optimization model minimizes the overall inventory cost while meeting the same service level of 95% for all end-products. We take the optimal base-stock levels, and rerun the simulation to verify the analytical results.

FIG. 5 shows the comparison between the “as-is” and the “to-be” models, in terms of the overall inventory cost. (To protect proprietary information, the vertical axis in all figures is normalized with respect to the inventory cost of the “as-is” model, which is 100.) As expected, the inventory cost at the end-product level is virtually eliminated in the “to-be” model. (The cost shown is due to WIP; the cost due to finished goods is nil.) In contrast, the “as-is” model keeps a significant amount of end-product inventory. On the other hand, the amount of component inventory is higher in the “to-be” model, which is again expected, since the required service level of 95% is common to both models. Overall, the “to-be” model reduces the overall inventory cost by about 30%.

Note in the above study, both models use the same demand forecast, at the end-product level. The “to-be” model, however, can easily switch to forecasting demand directly at the component level. This will result in improved forecast accuracy, as we can take advantage of parts commonality, as each component is generally used in several end products. Hence, in our study of the second factor, we evaluate the effect of forecast accuracy through a sensitivity analysis. FIG. 6 shows the overall inventory cost associated with three different levels of forecast accuracy. The first two columns, repeat the comparison in the last figure; i.e., both (“as-is” and “to-be” models assume 30% forecast error (i.e., the coefficient of variation equals 0.3) at the end-product level; the next two columns represent improved forecast errors, at 20% and 30%, achieved by the “to-be” model through forecasting at the component level.

Our study of the third factor aims at analyzing the impact on inventory when the system supports a richer product set, in terms of product variety. The motivation is to support mass customization. In the Internet-based, direct-sales environment, for instance, the number of different product configurations that customers want to order can be significantly larger than what is currently supported in the build-to-stock environment. FIG. 7 shows the inventory costs: the four columns on the left correspond to the current product set (1×), with the first scenario (S1) being the “as-is” model, and the other three being the “to-be” model at the current (S2) and improved (S3, S4) forecast accuracy levels, respectively; the four columns on the right repeat these scenarios with a product set that is ten times as large in variety (10×), while maintaining the overall volume. (Also refer to Table 12 for a summary of all the different scenarios.)

Observe that as the product variety increases, a significantly higher level of inventory is required in the “as-is” model. This is because forecast accuracy will deteriorate when the end products proliferate (i.e., larger varieties at smaller volumes). On the other hand, in the “to-be” environment, the increase in inventory cost is very modest. This is because the proliferation of end-products will have minimal effect on the forecast accuracy at the building-block level, due to parts commonality. This strongly supports the fact that the building-block model is the right process to support an internet-based, direct-sales operation.

TABLE 12
Summary of the Scenarios Used to Study
the Effect of Product Variety
Sce-
nario	Description	1x Cases	10x Cases
S1	“as-is”	original product set,	ten times larger product set,
		30% forecast error,	30% ×{square root over (10)} forecast error at
		90% service	MTM level
S2	“to-be”	forecast at MTM	ten times larger product set,
		level, 30% forecast	forecast error as in S2(lx)
		error, 90% service
S3	“to-be”	forecast at BB level,	ten times larger product set,
		20% forecast error,	forecast error as in S3(lx)
		90% service
S4	“to-be”	forecast at BB level,	ten times larger product set,
		10% forecast error,	forecast error as in S4(lx)
		90% service

The examples given demonstrate the advantages of the invention applied to the business of manufacturing personal computers (PCs). However, the invention has applications to other configure-to-order (CTO) manufacturing systems and is not limited to manufacturing PCs. Thus, while the invention has been described in terms of a single preferred embodiment, those skilled in the art will recognize that the invention can be practiced with modification within the spirit and scope of the appended claims.

Claims

1. A method of managing manufacturing logistics of multiple end products for which no finished goods inventory is kept and wherein the end products have a demand forecast that changes weekly, and while a manufacturing process migrates from an existing operation to a configure-to-order (CTO) operation, wherein a CTO operation is a hybrid of make-to-stock and make-to-order operations; comprising the following steps which are performed without keeping finished goods inventory for any end product; and each of which method step is performed while the manufacturing process migrates from the existing operation to the CTO operation: maintaining an inventory of components, which components, termed “building blocks”, are built to stock, each said component having a cost wherein the cost is stored in a database;according to the CTO operation which is the hybrid of make-to-stock and make-to-order operations, configuring-to-order multiple different end products using said components as defined by a bill of materials (BOM) structure wherein the BOM structure is stored in a database, wherein each end product involves at least one unique component; wherein the multiple different end products comprise six different end products;establishing a base-stock level for each of said components;computing the base-stock levels from a greedy algorithm which iteratively reduces inventory budget until a budget constraint is satisfied;performing an optimization run at a week's beginning and then updating base-stock levels weekly based on outcome of the optimization; andreplenishing said components from suppliers in accordance with said base-stock levels and using optimization processing so as to reduce a total cost of inventory of said components, with the optimization processing performed by a computer processor after having received input from the databases,wherein said cost of at least one component differs from said cost of at least one other component, and wherein said difference determines the result of said replenishing step;meanwhile taking customer orders from the Internet as the manufacturing process migrates from the existing operation to the CTO operation, wherein the CTO operation is a Web-based configure-to-order (CTO) operation where customer orders are being taken from the Internet.

2. The method of claim 1, including the processor receiving input from a database that contains optimization control parameters comprising budget limit on expected overall inventory cost and step size for a greedy algorithm, and also from a database that contains manufacturing system parameters.

3. The method of claim 2, in which the processor accesses the databases and first performs preprocessing steps, after which the processor performs optimization processing, after which the processor enters a postprocessing phase and performs a step of computing reorder points.

4. The method of claim 3, including after the computing of reorder points, translating the reorder points into days-of-supply and outputting to an output device or database days-of-supply for each component.

5. The method of claim 1, in which the steps are performed where the end products are a family comprising six different end products configured from a set comprising 17 different components which are obtained from external suppliers with different supply leadtimes.

6. A method of managing manufacturing logistics of multiple different end products for which no finished goods inventory is kept, wherein the multiple different end products are multiple different personal computers (PCs); and while a manufacturing process migrates from an existing operation to a configure-to-order (CTO) operation, wherein a CTO operation is a hybrid of make-to-stock and make-to-order operations; comprising the following steps which are performed without keeping finished goods inventory for any end product; and each of which method step is performed while the manufacturing process migrates from the existing operation to the CTO operation: maintaining an inventory of computers, wherein the components are stock computer component, which components, termed “building blocks”, are built to stock, each said component having a cost wherein the cost is stored in a database;according to the CTO operation which is the hybrid of make-to-stock and make-to-order operations, configuring-to-order multiple different end products using said components as defined by a bill of materials (BOM) structure wherein the BOM structure is stored in a database, wherein each end product involves at least one unique component; wherein the multiple different end products comprise six different end products;establishing a base-stock level for each of said components; andreplenishing said components from suppliers in accordance with said base-stock levels and using optimization processing so as to reduce a total cost of inventory of said components, with the optimization processing performed by a computer processor after having received input from the databases,wherein said cost of at least one component differs from said cost of at least one other component, and wherein said difference determines the result of said replenishing step;meanwhile taking customer orders from the Internet as the manufacturing process migrates from the existing operation to the CTO operation, wherein the CTO operation is a Web-based configure-to-order (CTO) operation where customer orders are being taken from the Internet.

7. The method of managing manufacturing logistics of end products recited in claim 6, including deriving the base-stock levels from a greedy algorithm which iteratively reduces inventory budget until a budget constraint is satisfied.

8. The method of claim 6, in which the end products have a demand forecast that changes weekly, and the method comprises optimization run at a week's beginning and then updating base-stock levels weekly based on outcome of the optimization.