A novel flexible model for lot sizing and scheduling with non-triangular, period overlapping and carryover setups in different machine configurations

This paper develops and tests an efficient mixed integer programming model for capacitated lot sizing and scheduling with non-triangular and sequence-dependent setup times and costs incorporating all necessary features of setup carryover and overlapping on different machine configurations. The model’s formulation is based on the asymmetric travelling salesman problem and allows multiple lots of a product within a period. The model conserves the setup state when no product is being processed over successive periods, allows starting a setup in a period and ending it in the next period, permits ending a setup in a period and starting production in the next period(s), and enforces a minimum lot size over multiple periods. This new comprehensive model thus relaxes all limitations of physical separation between the periods. The model is first developed for a single machine and then extended to other machine configurations, including parallel machines and flexible flow lines. Computational tests demonstrate the flexibility and comprehensiveness of the proposed models.


Introduction
The classic Capacitated Lot Sizing Problem (CLSP) does not sequence or schedule products within a period (Bitran and Yanasse 1982;Haase 1996;Karimi et al. 2003). In addition, it does not allow a setup to be carried over from one period to the & Masoumeh Mahdieh ma_mahdie@yahoo.com 1 Faculty of Industrial and System Engineering, Isfahan University of Technology, Isfahan, Iran next, even when the last product in a period is the same as the first product in the next period. Gopalakrishnan et al. (1995) developed a modelling framework for formulating the CLSP with setup carry over by introducing additional binary variables, and later incorporated sequence-independent and product-dependent setup times and costs (Gopalakrishnan 2000). Different studies have demonstrated that considering the setup carry-over significantly saves costs by decreasing the number of setups and releasing production capacity (Gopalakrishnan et al. 2001;Gupta and Magnusson 2005;Porkka et al. 2003;Sox and Gao 1999). This problem also called the capacitated lot sizing problem with linked lot sizes (Suerie and Stadtler 2003). A further issue for capacitated lot sizing is to determine a sequence for all products within a time period if setup times or costs are sequence-dependent. The CLSP is called large bucket problem since several item can be produced per period (Eppen and Martin 1987). Subdividing the (macro-) periods of CLSP into several (micro-) periods leads to discrete lotsizing and scheduling problem (DLSP) which is called a small bucket problem (Fleischmann 1990;Salomon 1991;Salomon et al. 1991Salomon et al. , 1997. The main serious restriction of the DLSP as a small-bucket formulation is not allowing both setup time and production time within a period. Thus this article focuses on the CLSP as a big-bucket formulation which is more flexible for integrating lot sizing and sequencing decisions. The CLSP partitions the planning horizon into a number of lengthy time periods, allowing setups of several products within the same period (a ''big bucket''). Gupta and Magnusson (2005) classified the CLSP literature according to extensions on sequence dependency of setup costs and times. They extended the framework proposed by Gopalakrishnan (2000) to include sequence-dependent setup times and costs. Haase (1996) modelled the Capacitated Lot sizing problem with Sequence-Dependent setup costs (CLSD) and included setup times (Haase and Kimms 2000) by assuming predetermined efficient production sequences and null inventory for the production of an item in a period. The general lot sizing and scheduling problem (GLSP) (Fleischmann and Meyr 1997) is very close to the CLSD but is more flexible since it eliminates the restrictions of the CLSD. Meyr (2000) included sequence-dependent setup times, resulting in the GLSPST and extended it to become the GLSPPL for parallel machines (Meyr 2002).
In their recent well-structured review paper, Copil et al. (2016) presented the historical development of the body of knowledge for simultaneous lotsizing and scheduling problem and discussed the recent trends. The GLSP has been known as the most flexible lotsizing and scheduling formulation in large buckets for representing different environments under slight modifications (Koçlar 2005;Koçlar and Süral 2005). Moreover, the need for only triangular setups is relaxed in the GLSP as it allows multiple lots of a product in a period as long as the lots of all products do not exceed the number of micro-periods in a period. Non-triangular setup times can happen in many industries such chemicals, food, beverages and oil. For example, in the animal-feed industry, some product families can cause contamination of other families so mixing equipment must be cleaned in order to avoid it. Cleaning can result in substantial setups that consuming scarce production A novel flexible model for lot sizing and scheduling with… 885 time. The amount of cleaning can often be minimised by producing an intermediate cleansing or shortcut product which can give rise to non-triangular setup times. In an alternative approach to the GLSP, Clark and Clark (2000) designed a mixed integer programming (MIP) model for the simultaneous sequencing and sizing of production lots on a set of parallel machines. They assumed non-triangular sequence-dependent setup times, no setup costs and the possibility of backlogging demand.
The problem of sequencing a set of lots with sequence dependent setups is related to the travelling salesman problem (TSP) and the vehicle routing problem (VRP) (Laporte 1992a, b). Almada-Lobo et al. (2007) presented two models for the CLSP with sequence-dependent and triangular setup times and costs using the Miller-Tucker-Zemlin (MTZ) subtour prohibition constraints (Desrochers and Laporte 1991). The main restriction of conventional TSP based models is permitting the production of only one lot per product per period which may well not be optimal when non-triangular setups exist. Clark et al. (2010) formulated a sequencing and lotsizing model with non-triangular setup times based on the asymmetric travelling salesman problem (ATSP) at an animal-feed plant. To solve the model, optimal solution methods based on iterative subtour elimination and patching were developed. In the ATSP-based models (Almada-Lobo et al. 2007;Clark et al. 2010), at most one lot per product can be produced in a period (and no subtour is permitted), so in the case of non-triangular setup, any optimal multiple production of a shortcut product is not allowed. Menezes et al. (2011) relaxed this restriction and allowed production of multiple lots per period (and correctly including connected subtours) by using an iterative model and method based on a potentially exponentially number of subtour elimination constraints (to exclude disconnected subtours). Clark et al. (2014) presented a stronger formulation than Menezes et al. (2011) for modelling the production of multiple lots of a product per period by using a polynomial number of multi-commodity-flow-type constraints (Claus 1984) to exclude disconnected subtours while allowing ones connected to the main sequence. Guimaraes et al. (2014) proposed a two-dimensional framework to classify the discrete time modeling approaches for lotsizing and scheduling problem. They also present a new formulation using commodity flow based subtour elimination constraints for the problem.
Setup overlapping has been studied by Suerie (2006) for small-bucket and by Sung and Maravelias (2008) for big-bucket formulations, but with sequenceindependent setup times and costs. Belo-Filho et al. (2013) extended the model by Suerie (2006) for small-bucket and proposed two models for the capacitated lotsizing problem with backlogging and setup carryover and crossover. Almada-Lobo et al. (2007) incorporated setup carryover features for a capacitated lot sizing and scheduling problem that allows a product to be set up at the end of one period and the actual production to start in the next period. Menezes et al. (2011) modelled setup cross-overs that allows a setup to start in one period and to end in the next period.
In this article, the first mixed integer linear programming formulation is presented for lot sizing and scheduling with non-triangular sequence-dependent setup times and costs that allows not only multiple lots of a product in a period using just a polynomial number of constraints and incorporating all the necessary features of setup carryover, as in Clark et al. (2014), but also overlapping of setups over period boundaries. The inclusion of overlapping setups is the original contribution of this article and permits modelling the production system more realistically by relaxing all the limitations of physical separation between the periods.
Moving towards more flexible and realistic modeling in production planning systems has been already attracted many researchers. To alleviate the problem of physical separation in discrete time scale, an alternative approach called block planning is proposed based on continuous representation of time (Günther 2014;Günther et al. 2006). However, the degree of flexibility of proposed approach is limited to necessity of the the grouping of product into setup families and the production of product within a family in a pre-defined sequence.
For the first time, in this paper not only all the limitations of discrete time scale modeling are relaxed but also practical assumptions are researched.Thus, a setup can start at the end of a period and finish at the beginning of the next period, or a setup can finish at the end of a period and production start in the next period. Furthermore, an imposed minimum lot size can cross over periods, and the setup state is conserved when no product is being processed over multiple periods. All these features increase the model flexibility and lead to better solutions, particularly under tight capacity conditions or whenever setup times are significant. The extension of the model to parallel machines or a flexible flow line is presented and discussed via computational tests.
The new model for single machine is developed in Sect. 2, allowing the production of multiple lots while incorporating all the features of setup carryover and overlapping. Moreover the effectiveness of multi-lot over single-lot production by taking advantage of shortcut products and the usefulness of modelling the setup overlapping under tight production capacity are both illustrated in some examples in Sect. 2 and then computationally tested in Sect. 3. The model is extended to parallel machines and flexible flow lines in Sect. 4 where the efficiency of each model is discussed in detail with an example. The paper concludes in Sect. 5 with a discussion of the model's value and identifies remaining challenges and opportunities for future research.

st ij
Time needed to setup from product i to product j. sc ij Cost of setting up from product i to product j. b i Time needed to produce a unit of product i. h it Cost of holding a unit of product i in inventory from period t to t ? 1. g it Backlog cost per period for product i from period t to t ? 1. UB it Upper bound C t =b i on the quantity of product i produced in period t. i 0 The product that is already setup at the end of period 0, i.e., the starting setup configuration in period 1. ml j Minimum lot size imposed on product j.
The decisions made by the model are represented by following variables: I it Inventory level of product i at the end of period t. B it Backordered amount of product i at the end of period t.
x it Production quantity of product i in period t. Slk t Number of units of slack capacity in period t. x F it The quantity produced in period t of the first (crossover) lot of product i in period t if it was setup in period t -1, otherwise 0.
The quantity produced in period t of the last (crossover) lot of product i in period t if its production continues into period t ? 1, otherwise 0. y ijt Number of times that production is to be changed over from product i to product j in period t. Integer non-negative. z it Number of times that product i is in a setup state in period t, Integer nonnegative. a it =1 either because j-to-i is the last setup in previous periods to t or because jto-i is the setup operation that overlaps from t -1 to t.
For all the products, the initial inventory ðI i0 Þ and the backlogs ðB i0 Þ are set to be zero at the start of the planning horizon.

The objective function and main constraints
The objective function minimises a weighted sum of backorders, inventory and setup costs: Constraint (2) balances inventory, backlogs, production and demand over consecutive periods: Constraint (3) represents the limited capacity and calculates any slack capacity: Constraint (4) enforces the appropriate setup before production: x jt UB jt Â z jt 8 j; t Constraint (5) prohibits setup between the same products: Constraint (6) ensures that the machine is set up for exactly one product at the beginning of each period. The initial setup configuration at first period is expressed by constraint (7).
2.2 Imposing a minimum lot size Some cleansing products k require a minimum lot size ml k to eliminate the previous product's contaminants, and also prohibits that a setup from i to j passes through cleansing products k without any production. Constraints (8)-(11) achieve this and also allow a minimum lot size to cross over the periods.
Recall that x F jt is the quantity produced in period t of the first (crossover) lot of product j in period t if it was setup in period t -1, but is otherwise 0, as imposed by constraints (8): Similarly x L jt is the quantity produced in period t of the last (crossover) lot of product j in period t if its production continues into period t ? 1, otherwise 0, as imposed by constraints (9).
x L jt UB jt a j;tþ1 8 j; t Then x L jt þ x F j;tþ1 is the size of a crossover lot of a product j that has been started in period t and completed in period t ? 1. Constraints (10) oblige this crossover lot to be of size at least ml j : Lastly constraint (11) imposes minimum lot sizes for both crossover and noncrossover lots using auxiliary variables x L jt ; x F jt .
x jt À x F jt À x L jt ! ml j z jt À a jt À a j;tþ1 Constraints (11) force a lot to be of size at least z jt ml j in period t. If the machine begins or ends the period in setup state j (or both) then a jt þ a j;tþ1 ¼ 1 or 2 ð Þ then constraints (11) impose the z jt À a jt À a j;tþ1 À Á lots to be at least of size z jt ml j , splittable into smaller separate lots of at least size ml j units in size. Clark et al. (2014) imposed a minimum lot size with the condition that there exists at least one setup in each period, i.e., result a carryover lot could not span over whole periods. Letting a carryover lot span over 3 or more periods while forcing the minimum lot size for the whole crossover lot was left as a challenge for future research. In this paper, this limitation is removed. The following example shows how the new minimum lot constraints can span the lot over the periods with no demand and impose the minimum lot size ml j À Á for the whole crossover lot.
Example 1 Consider a demand for product A in period 1, for product B in period 3 and no demand in period 2. A minimum lot size is imposed on the use of shortcut product C. In this case there are two possibilities as now detailed below: In the first possibility, setup A to C and C to B can both happen either in period two or, one setup can happen in period two and the other setup in period 1 or 3. So the minimum lot size will be enforced by constraint (11). In the second possibility, setup A to C happens in period 1 and setup C to B in period 3 while there is no setup in period 2 as shown in Fig. 1. So according to constraint (10): and according to constraint (11): In order to impose the minimum lot size for C, it is necessary to justify that the total production of product C (at the end of period 1, in period 2 and at the beginning of period 3) is at least ml C : To justify this, first constraints C1 and C2 are summed: Then constraints C3, C4 and C5 are summed: Finally combining constraints C6 and C7 concludes that the crossover lot of product C x C1 þ x C2 þ x C3 ð Þ is at least ml C and constraint (10) imposes ml C (not 2ml C ) for the whole crossover lot. Moreover this conclusion can be extended for more than one period with having no demand.
Note that constraints (8)-(11) are more efficient than the conventional constraint: x jt ! ml j P i y ijt ; 8 j; t, as used in other lot sizing and scheduling models (Clark and Clark 2000;Fleischmann and Meyr 1997) to impose minimum lot size. The reason is that in the conventional constraint, the whole setup and the production of the minimum lot size should be carried out in a single period so the minimum lot size neither can crossover to the next period(s) nor can be produced in a period when the setup is ending at the end of previous period(s). All these restrictions are relaxed in the new constraints (8)-(11). Examples 2 and 3 in the Sect. 2.4 show explicitly the difference of two types of constraints for imposing minimum lot size.

Lot sequencing constraints
Here, the ATSP-related constraints are demonstrated for sequencing product lots. Conventional ATSP-based models restrict production to at most one lot per product per period, which may not be optimal when non-triangular setups exist. Nontriangular setups occur in industries such as food, animal feed, beverages and oil where there are intermediate ''cleaning'' or ''shortcut'' products. For example in the animal feed industry, some products can contaminate other products and lead to serious effects on animal's health. To avoid this, machines must be cleaned, sometimes resulting in substantial setups that consume scarce production time. Alternatively, the production of a sufficient amount of an intermediate or cleaning product can clean the machines and reduce overall setup times (costs). In this situation, the setup to and from the cleaning or shortcut product k ð Þ is less costly and time consuming than a direct setup between two products i; j ð Þ means that st i;j ! st i;k þ st k;j . Therefore the shortcut product may need to be produced more than once within a period.
A sequence with multiple lots per period for some products could look like that illustrated in Fig. 2. Subtours connected to the main sequence S by shortcut products are possible (such as subtours B and C). Thus an exact formulation must allow connected subtours but exclude disconnected subtours (such as subtours A and D). To model the sequencing of product lots, the multi-commodity-flow (MCF) formulations by Claus (1984) are adapted to exclude disconnected subtours while allowing ones connected to the main sequence. Clark et al. (2014) applied the Claus (1984) ATSP subtour elimination method to allow multiple productions of shortcut products for a single machine and computationally demonstrated the effectiveness of the Multiple-Lot (ML) model in comparison with the equivalent One-Lot (1L) models. In this work, the same method is applied and the constraints are as follows.
Constraints (12) and (13) are flow conservation constraints relating the a it and z it setup state variables to the y ijt changeover variables as shown in Fig. 3.
To make constraints (13) work for last period t ¼ T either set t ¼ 1; ::; T þ 1 f gis considered for a it or new constraints (13a) are added as follows: The optimal solution to the model specified so far is a sequence from product plus any disconnected subtours. The latter are excluded by imposing in every period t that there is so-called k-walk from  (12) and (13) ðij a it ¼ 1 f gÞto all products k in the period's sequence. From now on, p a t denotes product ij a it ¼ 1 f g: Define additional binary variable a k ijt as follows: a k ijt =1 if the arc i ! j is on a k-walk from crossover product p a t to product k within period t's sequence of lots, otherwise 0.
The arc i ! j has to exist, hence: Further binary decision variables z bin it are needed. Define: where ZUB i is a fixed upper bound (UB) on z it and greater than one. ZUB i can be estimated as the smaller of J (the number of products) and the size of the ordered set f i; j ð Þjst ij ! st ik þ st kj g, which is 1 for many non-shortcut products. Constraints (17-19) below exclude disconnected subtours. Constraints (17) force the k-walk to reach product k and are enforced only when the setup state k exists for a time in the period (i.e., when z bin kt ¼ 1), but not when this is never the case (when z bin kt ¼ 0): If there is no production of product k in a period, then z bin kt ¼ 0, and by (17), (14) also forces this via a k ikt y ikt ¼ 0). The k-walk corresponding to the variables fa k ijt j8i; jg has to begin at p a t and then pass through other products to reach product k.
If a kt ¼ 1 then there is no need for a k-walk. If a kt ¼ 0, then by (17) P i a k ikt ¼ 1, i.e., a k ikt ¼ 1 for precisely one product i, the penultimate on the k-walk. Then, by (18), a k jit ¼ 1 for precisely one product j that is the 3rd last product on the k-walk, and so on, reversing along the k-walk, requiring a k ijt ¼ 1 along the k-walk, finishing at the initially-setup product i ¼ p a t (for which a it ¼ 1).
Constraint (19) forces the k-walk from p a t to terminate at product k: If there is no production of k in period t, then (19) requires a k kjt ¼ 0 which is not constraining as a k ijt ¼ 0 by (17). The ML-SM model (Multiple Lot for Single Machine) is specified by expressions  (1-19). It allows multiple production lots of shortcut products for a single machine while still not relaxing the limitations of a period's physical separation.

Period overlapping setup constraints
The last step is allowing setup operations to overlap periods, i.e., to permit a setup to begin in a period and end in the next period. The model is called MLOV-SM and relaxes all limitations of physical separation between the periods. The MLOV-SM is advantageous when capacity is tight and so lot sizing and sequencing decisions need more flexibility to reduce backlogs.
Consider the following additional decision variables: OLS ijt =1 if the overlapping setup operation i to j begins in period t and finishes in period t ? 1, otherwise 0.

S t
The amount of setup time that overlaps into period t ? 1, having begun at the end of period t.
The value of S t must be zero if there is no overlapping last setup at the end of period t: The last setup and at most one setup in period t can overlap from period t to t ? 1: The value of OLS ijt must be zero if i to j is not a setup initiated in period t: The capacity constraint (3) now becomes: When the last setup is overlapping, OLS ijt ¼ 1, then product j cannot be produced as it is the last (crossover) lot in period t. Thus constraints (4) and (9) now become (24) and (25).

Examples
Two examples now show the effectiveness of the new minimum lot constraints (8) (1-19), and the Conventional Model which has the same constraints as ML-SM but imposes minimum lot sizes by conventional constraint (26) rather than new imposing minimum lot sizes constraints (8)-(11).
Note that in the ML-SM model, constraints (4) are valid but loose: the value of z jt need only be 1, and not C2. Thus constraints (4) can be tightened by replacing z jt by z bin jt x jt UB jt Â z bin jt .
Examples 2 and 3 The following data are used for both examples: the demands are shown in Table 1. The models are implemented in the optimisation modelling software GAMS build 24.7.1 (Brooke et al. 1988) and solved using the industrial-strength CPLEX 12.6 solver (CPLEX. 2014) on a computer with a 2.1 GHZ CPU and 2 GB of RAM. All models were solved in less than a second for both examples.
The production diagram and the results of Example 2 are shown in Fig. 4 and Table 2 respectively. Note how modelling of all necessary features of production improves the solution remarkably. As shown in Fig. 4, the Conventional model  cannot use the machine's capacity efficiently and there are 5 units of idle or slack time in period 1 as the setup and minimum lot production has to be done totally in a single period (constraint (26)). This restriction is relaxed in the ML-SM model so the setup ends in period 1 and the minimum-sized lot is produced in period 2 that significantly results in a reduction of the number of inventory and backlogs as shown in Table 2. However there are still 10 units of slack time in period 2 as, in the ML-SM model, a setup cannot overlap, i.e., the setup begins in period 2 and ends in period 3. In the new lot sizing and scheduling model, MLOV-SM, all the limitations caused by previous models are relaxed and the production system is modelled realistically. Thus the scarce production capacity is used more efficiently.
In Example 2 the optimal solution is obtained by the MLOV-SM model with no shortages or inventory. In order to tighten capacity even more, the demand of product 2 is increased to 95 in Example 3. The production diagram and the results of Example 3 are shown in Fig. 5 and Table 3 respectively. Note that the Conventional model found a solution with high total inventories (50) and backlogs (15) while the optimal solution found by MLOV-SM has no backlogs and only 5 inventories.
Furthermore, as shown in MLOV-SM's production diagram in Fig. 5, the minimum lot crosses over from period 1 to 2. Lot crossover is another feature which is modelled via the new minimum lot size (ml) constraints (8)-(11), improving the solutions and giving more flexibility to the lot sizing model. Examples 2 and 3 showed how the new comprehensive mathematical formulation, MLOV-SM, relaxes all limitations of physical separation between the periods. The MLOV-SM modelled the new features consisting of starting a setup in one period and ending it in the next period, ending a setup in a period and starting production in the next period(s), and crossing a minimum lot size over multiple periods.

Computational tests
The aim of the tests is to assess how effectively the Multiple Lot model took advantage of shortcut products to reduce the total time spent on setups, compared to the equivalent One Lot (1L) model. In the latter case, the formulation (ML-SM) can   be simplified to a model that assumes At Most One Lot per product per period (denoted 1L-SM) by merging z jt and z bin jt to be a binary variable z jt for a single machine. Thus constraints (15) and (16) disappear. The tests also evaluated the impact of model MLOV-SM, on reducing demand backlogs, total inventory and cost in the case of tight production capacity. The models were implemented in the optimisation modelling software GAMS build 24.7.1 (Brooke et al. 1988) and solved using the CPLEX 12.6 solver (CPLEX 2014) on a computer with a 2.1 GHZ CPU and 2 GB of RAM.
To obtain initial insights, the performance of the three models (1L-SM, ML-SM and MLOV-SM) was compared on two problem sizes: a small size with 10 products including 1 shortcut product, and a big size with 20 products including 2 shortcut products, whose lot sizes and sequences were to be scheduled over two horizons of T = 4 and T = 8 demand periods.
The following data were used: C t ¼ 100; ml j ¼ 5; i 0 ¼ 1; b j ¼ 0:5; h jt ¼ 10; g jt ¼ 10000; 8j; t for all instances. In Clark et al. (2014) the setup times were initially set to be st ij ¼ j À i ð Þ if j ! i otherwise 10 þ j À i ð Þ , so the product 2 would normally be setup immediately after product 1. However, product 5 was then made an extreme shortcut with zero setup times: st 5j ¼ st i5 ¼ 0: In this paper, to make setup times more tangible, particularly in case of an overlapping setup, all setup times were increased by 3 so that and for shortcut products are: sc 5j ¼ sc i5 ¼ 50: The periodic demand forecasts d it varied randomly over product i and period t to provoke non-uniform lot-sizes and avoid lot-for-lot production. To show the effectiveness of model MLOV-SM, the demands in two consecutive periods are set to be non-zero for different products for time horizon T = 4. For example, if there are 10 products, then for period t, 5 random products have non-zero demand, with the other 5 having demand zero, while in period t ? 1, those products with zerodemand in period t now have non-zero demand, with other 5 having zero demand. We also used another TBO-profile (time between orders) with different lengths 1, 2 and 3 for time horizon T = 8. In this case, for each product a random TBO length (from 1 to 3) is chosen and then demands are generated for a product over 8 periods according to the TBO.
When capacity is loose, then there is much more flexibility about when setups can occur in an optimal solution, so we expect that period-overlapping setups will not make a difference. However, under tight capacity, there will be little such flexibility, so it is important to use scarce production capacity efficiently via relaxing all restrictions of physical separation between the periods. To simulate tight capacity the overall demand was adjusted so that setup times could take up to 20-25% of capacity. For loose capacity this was adjusted to 15%.
A similar procedure was applied for big size problems with 20 products. The machine capacity per period was doubled and setup times for products P11 to P20 simply replicate those for P1 to P10, with the two extreme shortcut products being P5 and P15.
Considering the two types of capacity (loose and tight) and planning horizons (T = 4 and 8), 4 combinations were generated for each problem size. For each combination, 20 test problems were generated, totalling 160 problem instances for big and small sizes, which were solved by the 1L-SM, ML-SM and MLOV-SM models. The CPLEX optimizer was allowed to run for a maximum of 1 h for big size problems, at which point the incumbent solution (i.e., the best found up to then) was used. Table 4 compare the performance of three models on 6 criteria calculated over the planning horizons 4 and 8: Total time spent on setups = P ij st ij y ijt Amount of unused (slack) capacity = P t slk t For each criterion, the difference between the mean values for the three models was statistically tested using a balanced analysis of variance test. The test used the data instance (that is the run) as a random blocking factor. The null hypothesis is that the difference between the models' means is zero.
The results in Table 4 and the paired t test p values in Table 5 show a highly significant decrease in backlogs, inventory and total cost under tight capacity for the model MLOV-SM compared to those for the ML-SM and 1L-SM. It highlights how model MLOV-SM uses scarce machine capacity and how the relaxing of all restrictions of physical separation between the periods plays an important role in minimizing shortage. The ML-SM model is also more efficient than 1L-SM as it uses the shortcut product P5 in small size problem and products P5 and P15 in big size problems, to economise on setups and reduce backlogs and inventory.
As expected, under loose capacity with no backlogs, due to greater flexibility in setups, period overlapping did not make a significant difference in inventory and slack time, although it significantly improved the total cost compared to the 1L model.
Not surprisingly, there were much longer solution times for 20 products than 10 products, and also for instances with T = 8 periods compared to those with T = 4. For 20 products and T = 8 under tight capacity, 17 of the 20 instances of the MLOV model used the full 1 h allowance of computing time (with median optimality gap of 3.7% for these 17), while none did for the 1L and ML model.

Parallel machines
The input data required by the PM models are: Demand for product i realised at the end of period t. C mt Available capacity time of machine m in each period t. st ijm Time needed to setup from product i to product j on machine m. sc ijm Cost needed to setup from product i to product j on machine m. b im Time needed to produce a unit of product i on machine m. h it Cost of holding a unit of product i from period t to t ? 1. g it Backlog cost per period for product i from period t to t ? 1. UB imt Upper bound C mt =b im on the quantity of product i produced in period t on machine m. i 0m The product setup at the end of period 0 on machine m, i.e., the starting setup configuration.
The decisions variables by the PM model are represented by following variables: I it Inventory level of product i at the end of period t.

B it
Backordered amount of product i at the end of period t. x imt Production quantity of product i in period t on machine m.

Slk mt
Number of unites of slack capacity of machine m in period t. x F imt The quantity produced in period t of the first (crossover) lot of product i on machine m in period t if it was setup in period t -1, otherwise 0.
The quantity produced in period t of the last (crossover) lot of product i on machine m in period t if its production continues into period t ? 1, otherwise 0. y ijmt Number of times that production is to be changed over from product i to product j on machine m in period t, Integer non-negative. z imt Number of times that product i is in a setup state on machine m in period t, integer non-negative. a imt =1 either because j-to-i is the last setup of machine m in previous periods to t or because j-to-i is the setup operation that overlaps from t -1 to t. a k ijmt =1 if the arc i ! j is on a walk from crossover product p a t to product k within period t's sequence of lots on machine m, otherwise 0. z bin imt =1 if product i is ever in setup state on machine m in period t, otherwise 0. OLS imt =1 if the overlapping setup operation j-to-i on machine m begins in period t and finishes in period t ? 1.

S mt
The amount of setup time that overlaps into period t ? 1 on machine m, having begun at the end of period t.
For all the products, the initial inventory ðI i0 Þ and the backlogs ðB i0 Þ are set to be zero at the start of the planning horizon. All the ML-PM and MLOV-PM's constraints are similar to ML-SM and MLOV-SM respectively with the new adapted data and variables. The complete ML-PM and MLOV-PM models are presented in Appendices 2 and 3.
Example 4 Consider 2 machines in parallel. The aim is to satisfy the demand shown in Table 6 for 10 products over the 4 planning periods with minimal backorders, inventory and setup costs. The capacity of each machine is C mt ¼ 50, thus a total capacity of P m C mt ¼ 100 is available for each period. The remaining PM data is the same as for the SM problem: ml j ¼ 5; i 0m ¼ 1; b jm ¼ 0:5; h jt ¼ 10; g jt ¼ 10000; 8j; t. Also the setup times and costs of each machine replicate those for a single machine.
The production diagrams and the results obtained by solving the 1L-PM, ML-PM and MLOV-PM models are shown in Fig. 6 and Table 7 respectively. Note that in Table 7, the 1L-PM and ML-PM model found the solution with the same amount 7 of inventory, and amounts 6 and 2 of backlogs respectively, while the optimal solution found by MLOV-PM has no backlogs or inventory.
The solution is illustrated in Fig. 6, where each node or circle shows the product at the top and its lot size at the bottom, and each arrow demonstrates a setup and an overlapped setup in bold as below: product Lot size

Setup
Overlapped Setup Note in Fig. 6 how effectively the MLOV-PM model twice took advantage of overlapping setups on machine 1 to use machine capacity and reduce inventory, backlogs and slack time. Furthermore, both the multiple lot models, ML-PM and MLOV-PM, took advantage of shortcut product 5 to reduce the backlogs, compares to the one lot model 1L-PM.
x jm e t À I jEt þ B jEt ¼ d jt 8 j; t ð27Þ x jm e t À I jet ¼ X m eþ1 x jm eþ1 ;tþ1 8 j; t; e ¼ 1; . . .; E À 1 ð28Þ Constraints (27) and (28) express the material balance including backorders for end items and work in process respectively. Constraint (29) bounds backorders of end items in any period to be within a specified proportion of demand. This is the practiced assumptions in flexible flow shop manufacturing systems (Ö zdamar and Barbarosoglu 1999). Moreover the holding cost will be different at each stage so h it now becomes h iet which is the cost of holding a unit of product i from period t to t ? 1 at stage e. The complete models for Multiple Lots for Flexible Flow Lines, denoted ML-FFL, and Multiple Lots with Setup-Overlapping for Flexible Flow Lines, denoted MLOV-FFL, are presented in Appendices 4 and 5 respectively. Apart from the inventory balance constraints, the FFL's constraints are similar to PM's substituting index m with index m e .
Example 5 If the parallel machines production system is duplicated in series, then the result is a Flexible Flow Lines (FFL) production system with two stages in series and two parallel machines for each stage. In this case, the FFL data for each stage is exactly the same as for PM. The holding costs assume that successive stages add value, so that work-in-process holding costs will increase as material progresses along the line. To reflect this, a value-added percentage factor VAP is used, whose value is 1.2. The first stage's unit holding cost h it1 for product i is 10 and for the subsequent stages, h ite ¼ VAP Á h it;eÀ1 , e ! 2. Thus the second stage's unit holding cost h it2 for product i is h it2 ¼ 1:2 Â 10 ¼ 12.
To analyse the FFL in detail, it was solved by the three models 1L-FFL, ML-FFL and MLOV-FFL considering the demand of first and second period in Table 6. The production diagrams and the results of FFL for two periods are shown in Fig. 7 and Table 8 respectively.
In order to simplify the FFL production diagram, the one-period-backward shifted demand is considered for intermediate stages ðe\EÞ, meaning that x jm eþ1 ;tþ1 in the right hand of Eq. (28) changes to x jm eþ1 t . Thus for first stage, the inventory balance equation would be I j1;tÀ1 þ P m 1 x jm 1 t À I j1t ¼ P m 2 x jm 2 t ; 8j; t: Note that the ML-FFL model took advantage of shortcut products in both stages and efficiently used the capacity of all four machines to reduce inventory, backlogs and slack capacity, compared to the ML-FFL. As shown in Table 8, the backlogs and inventory fell to 2 and 0 respectively for the ML-FFL model, and both fell to 0 for the MLOV-FFL. Thus the MLOV-FFL used the total scarce production capacity of 4 machines more efficiently by taking advantage of overlapping setups three times (Fig. 7) and left no inventory, shortage and slack capacity.

Computational tests
To obtain some insight into the relative efficiencies of the three models in PM and FFL, a variety of problem sizes are solved in a 3-h time limit considering demand over 2 and 4 periods, similar to Table 6. The objective function and the CPLEX optimality gap of the models after every hour are shown in Table 9 for each problem size. The test results in Table 9 show, for all problem sizes, that the MLOV model obtains a better solution than the ML and 1L models after 3 h and that ML is more efficient than 1L due to its use of the shortcut product. However in large instances, the models left large optimality gaps, particularly MLOV due to its extra binary variables OLS ijm e t for overlapping setups.
Note that for both time horizons of 2 and 4 periods, adding a stage or a machine significantly increases the optimality gap. Moreover CPLEX could not find a feasible solution for any problem with the attributes bigger than E ¼ 3; M ¼ 3 and T ¼ 4 within the 3-h time limit, emphasizing the need for an efficient heuristic solution procedure for large problems.

Final remarks
This paper presented new mix integer programming formulations for capacitated lot sizing and scheduling with non-triangular sequence-dependent setup times and costs, incorporating all the necessary features of setup carryover and overlapping on different machine configurations. These features relax all limitations of physical separation between the periods provide more flexibility to the lot sizing model.
To assess how effectively the multiple lot model with setup overlapping took advantage of shortcut products and setup overlapping features to reduce backlogs and inventory, three models 1L, ML and MLOV were compared for three production systems SM, PM and FFL. The computational results showed that the multiple-lots and setup overlapping features of the model enable more efficient production than when the formulation excludes setup overlapping or is restricted to single lot per product per product.
On a single machine the results showed highly significant decreases in backlogs, inventory and total costs for the MLOV-SM model compared to those for the ML-SM and 1L-SM models. Furthermore ML-SM is more efficient than 1L-SM due to its use of the shortcut product 5 to economise on setups and reduce backlogs and inventory.
The tests on the PM and FFL models also confirmed the effectiveness of the new formulation. However, because of the increased number of binary variables in large instances, CPLEX exhausted the available RAM before terminating the branch-&cut search and leaving a large optimality gap.
To sum up, the test results above, although merely probing, and not conclusive, indicate that for all machine configurations the MLOV model obtains a better solution. Due to the importance of the number of binary variables in large instances, future research needs to develop efficient solution methods for different machine Table 9 Results of the three models for different problem sizes in PM and FFL systems  configurations. Future work will also computationally compare different demand data patterns with variables sizes on the SM, PM and FFL models. The High Multiplicity Travelling Salesman Problem (HMATSP) is a special type of the classical travelling salesman problem in which each node is visited multiple times. Sarin et al. (2011) incorporated the HMATSP model as a substructure to formulate lot-sizing problem involving parallel machines and sequence-dependent setup costs, also known as the Chesapeake Problem. The HMATSP can also be applied for scheduling family products with several identical items to be produced separately on a single machine. Modelling Multiple-Lot production per period based on the HMATSP formulations poses a very interesting challenge for future research.
While the multi-commodity flow (MCF) subtour elimination constraints do provide much tighter formulations, it is recognised that their inclusion can be increase computational time in larger-sized models. The challenge of improving computing times is left for future research.
Given that in the case of existing non-triangular setups sufficient production of an intermediate or cleaning product can clean the machine more efficiently, the question arises as to whether the quantity of cleaning product called minimum lot size is sequence dependent. This poses another research challenge about how to model the sequence-dependency of minimum lot sizes in lot sizing and scheduling problems.
x L jt UB jt a j;tþ1 À X i OLS ijt ! 8 j; t ð38Þ x L jt þ x F j;tþ1 ! ml j a j;tþ1 8 j; t ð39Þ x jt À x F jt À x L jt ! ml j z jt À a jt À a j;tþ1 X i b im x imt þ X ij st ijm y ijmt þ S m;tÀ1 À S mt þ slk mt ¼ C mt 8 m; t ð73Þ x jmt UB jmt Â z jmt À X i OLS ijmt ! 8 j; m; t ð74Þ y jjmt ¼ 0 8 j; m; t ð75Þ X i a imt ¼ 1 8 m; t ¼ 1; . . .; T þ 1 ð76Þ a i om mt ¼ 1 8 m; t ¼ 1 ð77Þ x F jmt UB jmt a jmt 8 j; m; t ð78Þ x L jmt UB jt a jm;tþ1 À X i OLS ijmt ! 8 j; m; t ð79Þ x L jmt þ x F jm;tþ1 ! ml j a jm;tþ1 8 j; m; t ð80Þ x jmt À x F jmt À x L jmt ! ml j z jmt À a jmt À a jm;tþ1 À Á 8 j; m; t ð81Þ a imt þ X j y jimt ¼ z imt 8 i; m; t ð82Þ X i b im e x im e t þ X ij st ijm e y ijm e t þ slk m e t ¼ C m e t 8 e; m; t ð97Þ x jm e t UB jm e t Â z bin jm e t 8 j; e; m; t ð98Þ y jjm e t ¼ 0 8 j; e; m; t ð99Þ X i a im e t ¼ 1 8 e; m; t ¼ 1; . . .; T þ 1 ð100Þ a i ome m e t ¼ 1 8 e; m; t ¼ 1 ð101Þ x F jm e t UB jm e t a jm e t 8 j; e; m; t ð102Þ x L jm e t UB jm e t a jm e ;tþ1 8 j; e; m; t ð103Þ x L jm e t þ x F jm e ;tþ1 ! ml j a m e j;tþ1 8 j; e; m; t ð104Þ x jm e t À x F jm e t À x L jm e t ! ml j z jm e t À a jm e t À a jm e ;tþ1 À Á 8 j; e; m; t ð105Þ a im e t þ X j y jim e t ¼ z im e t 8 i; e; m; t ð106Þ X j y ijm e t þ a im e ;tþ1 ¼ z im e t 8 i; e; m; t OLS ijm e t y ijm e t 8 i; j; e; m; t ð137Þ x jm e t ; I jet ; B jet ; S m e t ; slk m e t ; x F jm e t ; x L jm e t Positive variables z im e t ; y ijm e t Integer variables a k ijm e t ; a im e t ; z bin im e t ; OLS ijm e t Binary variables