A fix-and-optimize approach for the multi-level capacitated lot sizing problem

https://doi.org/10.1016/j.ijpe.2009.08.022Get rights and content

Abstract

This paper presents an optimization-based solution approach for the dynamic multi-level capacitated lot sizing problem (MLCLSP) with positive lead times. The key idea is to solve a series of mixed-integer programs in an iterative fix-and-optimize algorithm. Each of these programs is optimized over all real-valued variables, but only a small subset of binary setup variables. The remaining binary setup variables are tentatively fixed to values determined in previous iterations. The resulting algorithm is transparent, flexible, accurate and relatively fast. Its solution quality outperforms those of the approaches by Tempelmeier/Derstroff and by Stadtler.

Introduction

This paper treats the problem to determine time-phased production quantities (lot sizes) for multi-level production systems in which a changeover at a resource from one product type to another requires a setup time and/or causes setup cost. The capacity of the resources is limited. The time-phased demand is assumed to be given and has to be satisfied. To date, this type of lot sizing problem cannot be solved satisfactorily within computerized Material Requirements Planning (MRP) modules of Enterprise Resource Planning (ERP) systems. The reason is that these systems often ignore the capacity limits of the production system while computing the lot sizes. This leads to infeasible production schedules which result in long and unpredictable lead times and large in-process inventories.

The problem of finding capacity-feasible production quantities for a multi-stage production system that minimize setup and holding cost has been formally stated as the multi-level capacitated lot sizing problem (MLCLSP), see Billington et al. (1983). For capacitated production systems with non-zero setup times, the question whether a feasible production plan (without overtime) exists has shown to be already NP-complete, see Maes et al. (1991). Several authors have therefore developed heuristics for the problem. Multi-level capacitated lot sizing is hence both practically important and scientifically challenging.

From a practical point of view, generic lot sizing approaches for ERP systems should meet several requirements. First, they should be adaptable to specific aspects of a concrete production system. This favors flexible approaches based on mathematical programming as opposed to “hard-wired” procedural heuristics that follow a highly problem-specific logic. Second, they should enable the planner to find feasible solutions of acceptable quality quickly and then let him decide to invest additional computation time to improve the solution. Third, if lot sizing on the one hand and sequencing/scheduling on the other hand are treated separately, it should always be possible to disaggregate the (aggregated) lot sizes into a detailed production schedule in continuous time. In a multi-level production system, this can only be guaranteed if positive lead times between the production stages are considered. Assume a serial product structure with one end product that requires a single component. Both items are produced on different resources. Furthermore, assume that the resource for the end product is fully loaded to produce a single unit of the end product in a period. In this case, it is not feasible to produce a unit of the component in the same period so that it enters the unit of the end product at the beginning of the same period. Therefore, this unit of the component must be produced at least one period ahead. For this reason, a positive lead time is required.

Reviews of the literature on dynamic lot sizing are given by Bahl et al. (1987), Maes and van Wassenhove (1988), Gupta and Keung (1990), Salomon et al. (1991) and Buschkühl et al. (2008). Kuik et al. (1994) relate lot sizing to batching and comment on some general criticism of batching analysis. The progress with single-item lot sizing is analyzed by Wolsey (1995) and Brahimi et al. (2006). Drexl and Kimms (1997) discuss models that consider both lot sizing and scheduling. Karimi et al. (2003) give a review of solution approaches for single-stage capacitated lot sizing problems and Jans and Degraeve (2007) focus on metaheuristics for dynamic lot sizing. Quadt and Kuhn (2008) review capacitated lotsizing problems with extensions. Table 1 classifies papers on the MLCLSP by the general solution approach, i.e. mathematical programming-based approaches, Lagrangean relaxation and decomposition, decomposition and aggregation, local search and metaheuristics and finally greedy heuristics. Some of these papers are discussed below in more detail.

In Tempelmeier and Helber (1994), Helber (1995) and Tempelmeier and Derstroff (1996), early product-oriented decomposition approaches for the MLCLSP are presented. The Tempelmeier/Destroff-heuristic (TDH) is until now the fastest available method for the MLCLSP. It is based on a Lagrangean relaxation of the MLCLSP which then decomposes into single-product uncapacitated lot sizing problems of the type studied by Wagner and Whitin (1958). The Lagrangean relaxation leads to a lower bound on the objective function value. A heuristic finite scheduling procedure is used to create a feasible solution and to compute an upper bound on the optimal objective function value. While the algorithm is fast, it is difficult to describe and implement as well as inflexible with respect to modifications of the underlying problem. The solution quality especially for large problem instances offers opportunities for improvement. Katok et al. (1998) present a linear-programming (LP)-based approach that works with a heuristic modification of the coefficients of the production quantities in both the objective function and the constraints. Tempelmeier (2008, p. 342) shows that this concept is very vulnerable when setup times occur and capacity limits are tight. In these situations existing feasible solutions (without overtime) are not found.

Stadtler (2003) proposes a mixed-integer programming-based heuristic that solves a series of subproblems through internally rolling schedules with time windows. For the periods within the time window of a particular subproblem, the simple plant location variant of the lot sizing problem is used to speed up the optimization process. Stadtler's approach delivers high-quality solutions for problems with zero lead times, but cannot deal with positive lead times (Stadtler, 2003, p. 501). The reason is that in a general product structure multiple relations with different lead times between end product quantities and time-phased resource requirements can occur. Hence, this makes a consistent disaggregation into a detailed production schedule in continuous time impossible. In addition, solution times for his approach increase substantially as the problem size (number of binary setup variables) increases. A variant of Stadtler's general approach of internally rolling schedules for the Capacitated Lot Sizing Problem with Linked Lot Sizes (Haase, 1994, Haase, 1998) is presented by Sürie and Stadtler (2003). It is based on an extended model formulation and valid inequalities to yield a tight formulation that speeds up the branch&bound process.

Belvaux and Wolsey (2001) show how to develop tight formulations for several special problem features occurring in practice. Rossi (2005) develops a time-oriented decomposition similar to the one by Stadtler where some of the setup variables are initially relaxed while others are iteratively fixed. Unfortunately, the author provides no direct comparison to the procedures by Stadtler and by Tempelmeier. Pitakaso et al. (2006) present a decomposition algorithm for the MLCLSP based on a limited subset of products and periods. Each problem in the decomposition is solved to optimality and a capacity reservation mechanism is used to reflect products and periods “outside” of the current problem. The decomposition itself is controlled by an “ant colony optimization” algorithm. The computation times that are necessary to find better results than with Stadtler's method appear to be quite high (20–30 min) and then the average improvement is small. Berretta and Rodrigues (2004) describe a memetic algorithm which is also population-based. However, they compare their results to those presented by Tempelmeier only for tiny problems with 40 binary variables. The picture is similar for a similar algorithm by Berretta et al. (2005): Relatively large deviations from optimal solutions occur already for problems with only 60 binary variables.

In this paper we present a mathematical programming-based algorithm for the MLCLSP that is flexible and transparent that allows to trade in solution time for solution quality and that can deal with positive lead times. In an iterative fix-and-optimize approach (Pochet and Wolsey, 2006, p. 113, call this “Exchange”), a sequence of mixed-integer programs (MIPs) is solved over all real-valued decision variables and a subset of the binary setup variables. The solution with respect to the binary variables is a fixed parameter for the next MIPs that optimize other binary variables. The optimization of the algebraic decision model is done within the MIP solver and therefore the approach is flexible. The user can trade in solution time for solution quality by deciding about the number of binary variables to be treated within a single MIP and about the number of iterations in which these are optimized and fixed again. Even large problem instances from the literature can be solved with a high solution quality within seconds or few minutes. We study both the cases of zero and of positive lead times and compare our results to those for the approaches by Tempelmeier and Derstroff (1996) and Stadtler (2003) in an extensive numerical study. While the method by Tempelmeier and Derstroff (1996) is extremely fast (and much faster than ours), our method yields a substantially higher solution quality. It outperforms the one presented by Stadtler (2003) with respect to both solution quality and computation time.

The remainder of this paper is organized as follows: in Section 2 we give a formal definition of the MLCLSP. The algorithm to solve the problem is presented in Section 3. Numerical results comparing our method to those developed by Tempelmeier and by Stadtler are reported in Section 4. The paper ends with some conclusions and suggestions for further research (Section 5).

Section snippets

Problem statement and model formulation

The objective of the MLCLSP is to determine production quantities Qkt and end-of-period inventory levels Ykt of product k in period t so that the sum of setup, holding and overtime cost is minimized. The demand dkt per product and period at each production stage is given and has to be satisfied. Whenever a product is produced during a period, a setup is required during this period which results in both setup cost and setup time. The setup state γkt{0,1} for product k in period t is assumed to

Basic idea of the fix-and-optimize heuristic

The standard formulation of the MLCLSP based on production and inventory quantities as stated in Section 2 typically leads for all but tiny problem instances to solution times of a MIP solver that are prohibitively large. The number of binary setup variables (|K|·|T|) determines most of the numerical effort, while the number of real-valued variables is of secondary importance. The basic idea of our proposal is therefore to solve in an iterative fashion a series of subproblems that are derived

Test sets and reference values

In order to evaluate the performance of our approach, we used and extended test instances without lead times that were developed and documented in detail by Stadtler and Sürie (2000) and used in Stadtler (2003). Out of their test sets, we used test sets A+, B+, C, D and E as described in Table 4. These test sets cover different product structures with 10–100 items and between 16 and 24 periods.

In the original test sets by Stadtler and Sürie (2000), lead times are not considered. However, as

Conclusions and outlook

We have presented a mathematical-programming-based approach to solve the MLCLSP with lead times. Modeling lead times is necessary, if one wants to be able to disaggregate the time-phased lot sizes into a detailed production schedule in continuous time. Our method is based on an iterative optimization of a series of subproblems. Each of them includes all the real-valued decision variables, but only a specific limited set of “free” binary variables. The approach is easy to describe and to

References (54)

  • H. Stadtler

    Mixed integer programming model formulations for dynamic multi-item multi-level capacitated lotsizing

    European Journal of Operational Research

    (1996)
  • H. Tempelmeier et al.

    A heuristic for dynamic multi-item multi-level capacitated lotsizing for general product structures

    European Journal of Operational Research

    (1994)
  • L.A. Wolsey

    Progress with single-item lot-sizing

    European Journal of Operational Research

    (1995)
  • J. Xie et al.

    Heuristic genetic algorithms for general capacitated lot-sizing problems

    Computers and Mathematics with Applications

    (2002)
  • H.C. Bahl et al.

    Determining lot sizes and resource requirements: a review

    Operations Research

    (1987)
  • G. Belvaux et al.

    bc-prod: a specialized branch-and-cut system for lot-sizing problems

    Management Science

    (2000)
  • G. Belvaux et al.

    Modelling practical lot-sizing problems as mixed-integer programs

    Management Science

    (2001)
  • R. Berretta et al.

    Metaheuristic approaches for the multilevel resource-constrained lot-sizing problem with setup and lead times

    Asia-Pacific Journal of Operational Research

    (2005)
  • P.J. Billington et al.

    Mathematical programming approaches to capacity-constrained MRP systems: review, formulation and problem reduction

    Management Science

    (1983)
  • P.J. Billington et al.

    Heuristics for multilevel lot-sizing with a bottleneck

    Management Science

    (1986)
  • F.F. Boctor et al.

    Heuristics for the n-product, m-stage, economic lot sizing and scheduling problem with dynamic demand

    International Journal of Production Research

    (2005)
  • Buschkühl, L., Sahling, F., Helber, S., Tempelmeier, H., 2008. Dynamic capacitated lot-sizing problems — a...
  • H. Chen et al.

    A Lagrangian relaxation approach for supply chain planning with order/setup costs and capacity constraints

    Journal of Systems Science and Systems Engineering

    (2003)
  • A.R. Clark et al.

    A heuristic for a resource-capacitated multi-stage lot-sizing problem with lead times

    Journal of the Operational Research Society

    (1995)
  • A. Federgruen et al.

    Progressive interval heuristics for multi-item capacitated lot-sizing problems

    Operations Research

    (2007)
  • Y.P. Gupta et al.

    A review of multi-stage lot-sizing models

    International Journal of Operations and Production Management

    (1990)
  • Gutierrez, E., Hernández, W., Süer, G.A., 2001. Genetic algorithms in capacitated lot sizing decisions. In: Proceedings...
  • Cited by (175)

    View all citing articles on Scopus
    View full text