Production, Manufacturing and LogisticsQuantile regression metamodeling: Toward improved responsiveness in the high-tech electronics manufacturing industry
Introduction
The high-tech electronics manufacturing industry faces a number of important challenges. The industry is dominated by consumer goods that are highly driven by customer satisfaction measures. Globalization has made the supply networks more complicated and difficult to manage, while shorter product life cycles and increasing product complexity make it harder for the technological developments to keep up with market demands. In such environments, on-time delivery is often noted as the most important factor in predicting service levels (Boyaci, Ray, 2006, Mönch, Fowler, Mason, 2013). Accurate quoting of delivery dates also has a significant impact on inventory control, scheduling, and other decision-making problems in manufacturing systems. Consequently, the ability to meet the stated delivery dates across the supply chain is essential to today’s highly competitive global market. The importance of this not only to final product manufacturers but also to equipment suppliers for the high-tech electronics industry is highlighted in Atan, de Kok, Dellaert, van Boxsl, and Janssen (2016). Further complicating the industry’s ability to obtain high customer service levels is the fact that consumer demand changes rapidly. Therefore, being able to predict delivery dates for future operating conditions of the system is as important as generating accurate estimates of delivery dates for current operating conditions.
An essential component of providing customers with accurate delivery dates for a given product (both under current and future operating conditions) is the ability to quickly generate accurate estimates of parameters of the product’s CT distribution. Here, we consider CT as a random variable representing the time required for a job or a lot to get through the entire manufacturing process (including processing, waiting, and transportation times).
While estimates of mean CT are relatively easy to obtain and so are often readily available, using them to generate estimates of delivery dates ignores inherent variability associated with the CT distribution, and can result in reduced on-time delivery and service to customers. Estimates of quantiles of the CT distribution, on the other hand, provide the decision maker with a complete picture of the CT behavior, allowing delivery-date quotes to be made with varying levels of confidence (and associated risk). For example, if the 0.9-quantile of the CT distribution is used in setting up the delivery dates, 90% of the orders will be delivered on time. There is a growing amount of work in the estimation of quantiles (Alexopoulos, Goldsman, Wilson, 2012, Bekki, Mackulak, Fowler, Nelson, 2010, Calvin, Nakayama, 2013, Chen, Kelton, 2006, Chen, Kelton, 2008, Jin, Fu, Xiong, 2003, Raatikainen, 1987, Raatikainen, 1990). The importance and relevance of the use and estimation of CT quantiles in manufacturing systems are highlighted by Ankenman et al. (2011), Bekki, Mackulak, Fowler, and Nelson (2010) and Tai, Pearn, and Lee (2012).
Across all manufacturing systems, a key driver of CT is throughput (TH). We define TH to be the overall production rate of the system as a percentage of the system’s capacity, and it can be controlled at the operations level by the rate of introduction of new jobs or lots (start rate). Several other controllable operational factors, such as product mix and various distributional parameters of time-between-failures, repair time, setup time, loading and unloading times, have also been identified in the literature as closely tied to CT (Meidan, Lerner, Rabinowitz, & Hassoun, 2011). Therefore, determining the relationship between the CT distribution and these factors is critical in developing accurate CT predictions. If analytical models were available to describe these relationships, they would provide unambiguous functions that can be used in the prediction of CT. However, the use of such analytical models is restricted by the fact that they are only applicable to simple systems that satisfy fairly limiting assumptions. Consequently, in the complex and stochastic environment of the high-tech manufacturing industry, discrete-event simulation models are typically used to predict CTs (Mönch, Fowler, & Mason, 2013).
Though regularly utilized, simulation models of high-tech electronics manufacturing facilities are quite complex, and can take a considerable amount of time to execute for a given set of input variable values. The strain of executing such models is particularly evident when what-if scenarios are examined for analyzing alternate or future operating conditions where input variables take various new values. Further elongating the time to perform such analyses is the fact that multiple simulation replications need to be run for each examined what-if scenario to mitigate the impact of stochasticity in simulation models. Instead, it would be desirable to develop a regression function approximating the simulation input–output relationship based on the current input variable values and use this regression model to predict how the system would perform at the new input variable values that are considered in the what-if scenarios. Such regression models of simulation input–output relationships are known as metamodels, and have become an increasingly popular modeling tool (Barton, 2009). Ideally, metamodels provide the fidelity of the full simulation model with the ease of use of a spreadsheet environment.
Metamodels are tools developed to be used in production planning (Ankenman et al., 2011). For example, a practitioner could investigate the impact to a CT quantile of reducing the coefficient-of-variation (CoV) of unloading times, perhaps through increased automation. Similarly, an operations manager could quantify the resulting impact to a CT quantile of increasing the production rate of a product that his facility produces in response to anticipated demand changes for the product. In both cases, this information can be used not only to evaluate the impact and significance of potential, future changes to the production system, but also to quickly adjust lead-time quotations to customers if the changes are ultimately implemented on the manufacturing floor. Some applications of simulation metamodeling in the analysis of production systems can be found in Kesen, Toksari, Güngor, and Güner (2009), MacDonald and Gunn (2011) and Yang (2010).
An important part of the metamodeling process is the selection of a metamodeling (regression) function, g(x), that allows the prediction of the desired output performance measure for a given vector of inputs, x, in a simulation model. The mean CT is the most commonly used output measure in the simulation metamodeling literature, and there are several papers that focus on the relationship between the mean CT and TH (Fowler, Park, Mackulak, Shunk, 2001, Park, Fowler, Mackulak, Keats, Carlyle, 2002, Veeger, Etman, van Herk, Rooda, 2010, Yang, 2010). However, as previously discussed, quantiles provide a more comprehensive understanding of the CT distribution and are of greater use to decision makers.
When the CT output variable is denoted by Y with distribution FY, the q-quantile is defined as We propose the use of the Quantile Regression (QR) method (Koenker, 2005) to construct a metamodeling function, g(x), that is suitable for predicting CT quantiles given a vector of inputs, x, in simulation models of manufacturing systems using the following relationship where the randomness of the simulation model is represented by the random error ε. A strength of the proposed method originates from the fact that no distributional assumptions are made for FY, and hence ε.
In summary, the goal of this work is to develop a computational framework for predicting CT quantiles of high-tech electronics manufacturing facilities. We will develop and demonstrate the QR metamodeling procedure, which is of particular use for predicting CT quantiles of manufacturing systems in which there is more than one input variable of interest. Although the theory of QR modeling is well developed, the application of it to CT quantiles in manufacturing systems has many theoretical and computational challenges. This paper addresses the issues related to specifying the form of the quantile metamodeling function g(x) in the case that multiple input variables are present and finding a parsimonious model fit is technically challenging. Notably, while the proposed approach can theoretically be used for a large number of input variables, the work presented here demonstrates the efficacy of the approach up to only three input variables.
There are only a small number of studies that focus on the relationship between CT quantiles and TH (Chen, Zhou, 2011, Yang, Ankenman, Nelson, 2008). Both of these techniques cannot be extended to metamodels of input variables other than TH. Our contribution in this paper is to consider multiple input variables simultaneously, not only TH, but also product mix and various distributional parameters of time-between-failures, repair time, setup time, loading and unloading times, etc. To the best of our knowledge, this is the first study that shows how QR modeling can be used to build metamodels of CT quantiles with multiple input variables in the manufacturing environment.
In the upcoming sections of the paper, we more specifically articulate the problem and describe the specific simulation model that was used as an experimental testbed for validating the proposed mechanism of predicting CT quantiles of high-tech electronics manufacturing facilities. We then provide the rationale for the proposed approach and give the details of the QR metamodeling procedure. Later in the paper, experimental results are provided, followed by discussions and future work.
Section snippets
Statement of the problem
The objective of this study is to provide a computational framework based on simulation metamodeling for estimating the CT distribution at varying values of input variables in a semiconductor manufacturing environment. We propose the use of the QR method for estimating the metamodeling function, and highlight that the approach handles multiple input variables. We assume the existence of a validated simulation model of the manufacturing system under study. Our method is general enough, in the
Prototype semiconductor manufacturing system
The Minifab model is a simplified model of a semiconductor manufacturing facility designed to capture the key characteristics that make the modeling and scheduling of semiconductor manufacturing processes particularly difficult: re-entrant flow, batching, setups, preventative maintenance, emergency maintenance, and multiple product types. While still capturing the most important system characteristics, the execution time for the Minifab model is much less than that of a full semiconductor wafer
A review on the Quantile Regression (QR) method
A metamodel is a regression function approximating the input–output relationship of the underlying simulation model. The major steps in the development of a metamodel are: (i) choosing a functional form for the metamodeling function, g; (ii) designing and executing the experiments to fit the metamodel (i.e., the selection of the set of design points at which to observe the output Y); and (iii) fitting the metamodel and validating the quality of its fit (Barton, 2009).
We utilize polynomial
Quantile Regression Metamodeling (QRM) procedure
The QRM procedure accommodates multiple predictor variables. We recommend that the first predictor variable is TH, and the other predictor variables represent other most important, controllable factors influencing the CT distribution. To successfully utilize this procedure, we assume that the user will handle initialization bias present in the model independent of the QRM procedure.
In the QRM procedure, Monte Carlo cross-validated mean absolute percentage error (MCCV MAPE) is used as the model
Monte Carlo cross-validated mean absolute percentage error (MCCV MAPE)
The shrinkage parameter used by the Lasso approach, λ, and the order of the polynomial form of the QR metamodel to fit, k, play a vital role in determining the resulting QR metamodel’s predictive accuracy. That is, based on the same simulation output data set, different combinations of (k, λ) lead to distinct QR metamodels whose predictive performances may be substantially different.
Since the true quantile values are not available to evaluate the predictive accuracy of the QR metamodels during
Sensitivity of the procedure to key experimental parameters
While MCCV MAPE is listed in the QRM procedure as the model selection metric, it is one of three different metrics that were investigated for effective metamodel selection: (i) the loss function (L), (ii) the Akaike Information Criterion (AIC) (Akaike, 1973), and (iii) the Monte Carlo cross-validated MAPE (MCCV MAPE) (Shao, 1993).
The loss function, L, for a given QR metamodel is given in Eq. (1) and can be obtained directly from the fitting process.
AIC provides a measure of the relative quality
Computational analysis
In an effort to demonstrate the capabilities of the QRM procedure, the Minifab model presented in Section 3 was used to generate metamodels for estimating quantiles. In the first set of experiments that we ran, there were two predictor variables. In the second set of experiments, there were three predictor variables. In all cases, the objective was to build a metamodel to estimate the 0.5, 0.8, and 0.95-quantiles of the CT distribution.
Discussion and future work
The QRM procedure presented here addresses the need for an accurate and efficient process that simulation modeling practitioners can use to generate estimates of cycle-time quantiles from high-tech electronics manufacturing systems. A simulation model of a simplified semiconductor manufacturing system was used to generate an experimental testbed. Results demonstrated that the QRM procedure generates accurate estimates of cycle-time quantiles at points that were not simulated to generate the
Acknowledgment
We would like to thank Sarah A. Valente for her help with the simulation modeling of our example problem. Also, we would like to thank Rishikesh Nimma and Santosh Vemula for their help writing and executing the code for calculating the MCCV MAPE values. The authors thank the editor and two anonymous referees for their constructive comments, which helped improve the paper significantly. The work of the third author is partially supported by the ICTAS Junior Faculty Award at Virginia Tech (No.
References (36)
- et al.
Simulation-based estimation of cycle time using quantile regression
IIE Transactions
(2011) - et al.
Building metamodels for quantile-based measures using sectioning
- et al.
A framework for analysis and production authorization card-controlled production systems
Production and Operations Management
(2011) Statistical predictor identification
Annals of the Institute of Statistical Mathematics
(1973)- et al.
A new perspective on batched quantile estimation
Proceedings of the 2012 winter simulation conference
(2012) - et al.
Accurate emergency department wait time prediction
Manufacturing & Service Operations Management
(2016) - et al.
Simulation in production planning: An overview with emphasis on recent developments in cycle time estimation
- et al.
Setting planned leadtimes in customer-order-driven assembly systems
Manufacturing & Service Operations Management
(2016) Simulation optimization using metamodels
- et al.
Steady-state quantile parameter estimation: Empirical comparison of stochastic kriging and quantile regression