1) Aggregate variables:

• denotes the demand for a family of products (i.e. the level of aggregate sales during period). It is assumed to be a stationary random variable, being approximated by a normal distribution function with mean and finite variance. The assumption of normality is based on the Law of Large Number [17] . Such a law can be interpreted here as follows: the sum of different patterns of probability distribution functions, which are related to different products of the same family, can be approximated by a normal distribution to represent this family. It is worth realizing that the probability distribution function of the demand usually depends on the type of the production process. For instance, considering a make-to-stock production process, where the evolution of demand is usually stationary over the periods, a normal distribution function, with mean and variance given, is statistically a good alternative to model the fluctuation of demand; see [18] for more details.

• denotes the amount of net inventory level at the beginning of period. This variable takes values from a dynamic balance system that depends linearly on the demand fluctuation. Based on the property of a normal stochastic process which says that the resulting linear transformation of a sequence of normal random variables is also a normal random variable (see [17] ) is possible to consider that the inventory variable follows a normal process with mean and variance.

• denotes the rate of aggregate production capacity during the period, being a decision variable to the problem. Note that if, for instance, the inventory balance system is running under a closed-loop control scheme, then the production rate depends on the net inventory level for each period, and, as a consequence, this rate must also be understood as being a random variable. Furthermore, if this dependence is linear, then the probability distribution function of the production variable will be similar and proportional to the distribution function of the inventory variable; see this feature of the stochastic process in [7] .

• denotes the amount of the regular workforce used in the period. This variable is assumed here independent of endogenous and exogenous factors that influence the environment of the company. Indeed, there is a set point level for regular workforce denoted here as. Thus, any excess in the levels of fluctuation of demand, involving an increase in the production rate, will be dealt with a policy of the use of temporary labor or overtime. This variable is assumed to be essentially deterministic.

• denotes workforce changes between the subsequent periods and. It is a deterministic decision variable that provides the number of employees to be included or removed from the workforce level between two adjacent periods.

2) Cost’s components

The total cost function of the HMMS model, denoted here as, is given by the sum of quadratic and linear functions whose components are described below:

• denotes the regular payroll cost, where and are constants used to adjust costs with labor.

• denotes the hiring and firing cost. Such cost is associated with a change in the size of the workforce from period to period. The constant can be used for asymmetry analysis in costs of hiring and layoffs.

• denotes the inventory and backorder cost. The component represents the optimal reference level for the inventory. Note that whenever the level of current inventory deviates from this target level, the cost associated tends to increase.

• represents the overtime cost, which depends both on the workforce and production levels for each period. Particularly, the term denotes the maximum amount of families of products that can be produced without overtimes.

Note that the manager must provide the constants. These estimates require time-consuming activities, such as statistical analysis, account information, and many managerial insights. In this way, a curve-fitting approach can be used to provide these coefficients; see [1] .

Based on the above notation, the classical HMMS model can be formulated mathematically as follows:

Note that the inventory balance equation given in the problem (1) is a stochastic process. Thus, the inventory variable is a random variable and, as a consequence, the model (1) represents a stochastic optimization problem. Therandomness explains the use of the mathematical expectation operator in the criterion. It is assumed that the initial inventory and workforce levels are known and given, respectively, by and. The level of is set equal to 0.

As can be noted from problem (1), the classical formulation of HMMS model does not explicitly take into account constraints on the decision variables; see also [1] . Indeed, any significant change in the levels of the decision variables will be penalized directly in the criterion. As a consequence, such criterion must be able to cover a sufficiently long span of time to ensure full recognition of the changes of these variables.

In this section, the HMMS model is modified in order to consider constraints on the inventory, production and workforce variables explicitly in the formulation. It is a consensus that the use of physical constraints explicitly in the model makes it more realistic for practical applications; see e.g. [16] and [19] . Based on this, the original formulation of problem (1) will be reformulated to include constraints on decision variables. Thus, assuming that it is possible to set lower and upper bounds for each decision variable of the model, and also to include a workforce balance constraint, the reformulated model is described as follows: finding a non-negative optimal

policy, which includes both production rates and workforce changes , to optimize the following production planning problem:

where and denote lower and upper bounds on the inventory variable; and denote lower and upper bounds of the production variable; and is the upper boundary related to the size of regular workforce plus overtime. Note that the parameter denotes the maximum amount of workforce change, which is estimated for each period k.

The optimal production policy is assumed to depend on the inventory

variable for each period of the planning horizon. Since is a random variable, the variable will also have a random behavior. Such dependence follows a mathematical structure that is given by the function. If has a linear structure, the probability distribution of the production variable will be similar to the probability distribution of the inventory variable. Thus, since the inventory variable is assumed here to be a normal distribution with known first and second statistics moments, it follows then that the production variable will also be normally distributed with known expectation and standard deviation

.

It is interesting to observe that the randomness of the inventory and production variables makes them probabilistic variables, and so to ensure that their physical boundaries are no violated, they must be taken as chance- constraints. Then, probabilistic indexes of inventory and production constraints, which are given respectively by and, are used to express the user’s expectation for the nonviolation of these constraints. These indexes vary within the interval, and each one of them has its specific practical interpretation. For example, the index is usually interpreted as the level of the customer satisfaction. Indeed, if the manager chooses close to 1 means that he wants to meet the demand for complete, whenever it occurs. For instance, for means that the manager expects to deliver the products on time at least 95% of the time. In order to guarantee a high level of customer satisfaction, the manager must adopt a safety stock policy that allows ready delivery. In short, varying in the range, it is possible to analyze different production scenarios based on safety stock policies, which imply in greater or smaller level of customer's satisfaction. Finally, note that the index is used here to represent the productivity degree of the production process. This index is related to the production capacity required at each period of the planning horizon. The idea is to guarantee that unexpected orders, which may occur during a given period, can be promptly attended; avoiding, thus, the occurrence of backorder; see [20] .

The constrained HMMS model (2) is placed into an equivalent state-space linear-quadratic Gaussian model. For such, the variables of the HMMS model are considered as state and control variables, that is, and denote the state variables, and and denote the control (or input) variables. Based on this variables, the transformation process can be performed as follows.

The time-discrete stochastic state-space system that represents the linear balance of inventory and workforce is given by:

where the first and the second row of (3) describe the inventory and workforce linear balance equations, respectively. Note that is a random variable that represents the uncertainty related to sales fluctuation over the periods of the planning horizon. As mentioned previously, sales fluctuation brings randomness to the inventory balance Equation (3). Thus, the inventory must also be seen as a random variable whose probabilistic distribution function is similar to the distribution of demand. Since is stationary and normally distributed,

the distribution function of the random variable will be also normal and exactly known from its mean and variance [10] .

The dimension of the state-space system (3) will be augmented by the introduction of a model that represents the fluctuation of demand. For this, it is assumed that all information about customer demand is recorded in the sales history set. Based on such information, an Auto-Regressive, Moving Average (ARMA) model is identified and can be used both to demand forecasting and to data generating process; see [12] .

Once a model has been identified, it is placed into space-state format and coupled to the system (3). The transformation process is carried out as follows: Firstly, the demand variable is decomposed, for each period, in two variables, that is. The first variable denotes the mean value of demand, and the second variable is the residual demand, i.e., stationary random variable with normal distribution and finite variance. As a result, the variable can be mathematically described by an model given by Equation (4):

where denotes the white noise and represents the delay operator (for instance,). Note that the parameters of the sequences and are respectively Auto-regressive and Moving- average parameters of an ARMA model.

After some algebraic handling, the model (4) is converted into an equivalent state-space format that is given by:

where represents the equivalent state vector related to residual demand. The matrices of the system (5) are defined as follows:

and

where.

It is worth adding that to convert the input-output model (4) in the state-space model (5), the reliability condition should be guaranteed, i.e.. Note that if, then; see [21] .

Finally, handling Equations (3), (4), and (5), it is possible to obtain a general state-space system similar to (1) that represents the production process as whole, that is:

where;; and denotes a vector of random variables with the mean and covariance matrix. The matrices and vectors of the system (6) are given, respectively, by:

; and

It is assumed that the inventory, production, workforce and workforce change variables must take values from their specific solution spaces. Since the first row of the Equation (3) is a stochastic process, it is impossible to guarantee a priori that both inventory variable and production variables will not be violating their respective physical boundaries. Consequently, chance-constraints on inventory and production variables must be considered in order to overcome such difficulty, see Section 2.2. Finally, the constraints of the model in the state-space format are given by:

where represents the safety stock level; denotes the maximum storage capacity allowed in the period; and are the minimum and maximum production capacity, respectively; and denotes the maximal level of workforce during a period.

The original cost components of the HMMS model, given by the criterion in (1), is placed now in a state-space format. Without any loss of generality, it is assumed that the constants and are both set equal to zero, see [22] . The main reason of this assumption is that these constants do not have any effect on the choice of production variables, and so they can be ignored. Furthermore, to preserve proportionality, during the transformation of costs into the state-space format, the constant is set equal to the product between and (i.e.,). Note that such artifice does not take away the originality of the costs of HMMS model. Indeed, for many practical applications, is relatively close to; see [1] . Based on the exposed above, the original costs of HMMS model (1) can be placed in a matrix format, as follows:

where

•

•

•

•

Note that the coefficients of the criterion (8) are calculated by the solution of the following system of equations:

with_, see [1] . Note that denotes the absolute mean of the expected demand over the planning horizon, that is,; see [17] .

At last, with the objective of finding production and workforce optimal sequence plan, described by, a multi-period, single-product, stochastic production planning problem, which includes an ARMA model and constraints on the variables, is formulated as follows:

The problem (9) is an extended version of the HMMS model, and it also belongs to the class of the linear, quadratic Gaussian model with constraints on state and control variables, see [13] and [16] . The main advantage is that it can be employed to model and optimize a wide variety of management problems in a unified format, see for instance [7] [9] [15] [19] [22] and [23] . The next section discusses the solution of the problem (9) by mean of two suboptimal approaches of the control theory, which are of simple implementation.

Having been assumed previously that the demand is normally distributed, then it is possible to conclude that the statistical behaviour of the linear inventory system given in (6) follows a normal process. This charac-

teristic means that the probability distribution function of inventory, given by, can be precisely computed

from (6) by mean of the determination of its respective mean and variance equations; see [20] . These two statistical moments allow transforming the stochastic problem (9) into a Mean Value Problem (MVP).

The first step for converting the stochastic problem (9) to an equivalent MVP is to determine the mean and variance of the state variables of the system (6). However, it is important to bear in mind that the inventory

variable is the only random variable of the vector, with mean and standard deviation. As a result, the mean vector and the covariance matrix of the system (6) are given as follows:

Note that the other state variables (i.e.,) of the system (6) are essentially deterministic, then only to guarantee the uniformity of the notation it will be considered here that

.

Since the control variable depends on the state variable, it can be completely defined by its mean and covariance. Note also that the mean and covariance of the demand variable are given respectively by and, with jÎ, as discussed previously in Section 2.3.2.

Based on these statistics, the criterion (8) and the linear system (6) can be promptly converted to a deterministic equivalent pattern [20] . Another important transformation is to convert the probabilistic constraints (7.a) and (7.b) in equivalent deterministic inequalities. Thus, from (7.a) follows that:

and,

where denotes the inverse probability distribution function of the inventory variable, which depends on the service level.

Proceeding in a similar way, follows that the probabilistic constraint (7.b) becomes:

where;; and denotes the inverse probabil-

ity distribution function of the production variable, which depends on the production capacity reliability level, i.e., the probabilistic index.

At last, it is important to observe that the state variables related to the ARMA model (i.e,)

are totally unconstrained. This feature means that such variables can evolve freely along the planning horizon without any control.

After all these transformations, the Mean Value Problem can be formulated as follows:

It is worth mentioning that the problem (14) allows not only to provide an optimal aggregate production plan, but also to help the manager to get important insights about the use of the aggregate resources. Indeed, from varying some parameters of the problem (14), it is possible to analyse different scenarios related to an inventory-production process. For example, comparing optimal inventory, production and workforce trajectories, developed from different scenarios analysis, the manager can realise how the future will be like in terms of the reorder cycle for a given product. The prior knowledge of all possible reorder points allows that some actions can be in advance taken to replenish the inventory levels, and, thus, to prevent against the possibility of stock out. These actions can be understood as an attempt to meet future demand for aggregate products, and also to prevent against unexpected events, such as, for instance, delays and machines breakdowns.

Two simple suboptimal approaches, known in the literature as Open-loop No-updating and Open-Loop Updating [13] , can be combined with quadratic programming algorithms for solving (14), and so generating a sequential optimal plan. These approaches are briefly described below.

• The Open-Loop No-updating (OLN) approach applied to (14) provides an optimal production policy that is entirely conditioned on initial information. Even if further information becomes available, these initially computed policies are enacted up to the end of the time horizons,” see [15] . This characteristic means that OLN procedure provides an optimal sequence that depends only on the initial state of the system (6), i.e.,. Thus, any available information on the state of the system is completely ignored for period. Consequently, the OLN approach does not take into account any feedback scheme in its strategy of providing a feasible solution for the problem (14).

• The Open-Loop Updating (OLU) approach is very similar to the OLN, except that optimal policies are always recomputed as soon as new information about the state of the system becomes available [15] . In fact, for each new period, whenever the current state of the system is measured, it is assumed to be a new-initial state to solve the problem (14) from the current period to the end-period. As a result, the optimal sequence is provided. From this sequence, only the first vector, i.e., is applied to the system (6), while the others are completely ignored. Note that this approach must be repeated N times in order to provide optimal control policy for the problem (14).

Both OLN and OLU approaches can be used to develop production scenarios. These scenarios are useful for managers getting insights on the use of the resources of the company. Also note that such scenarios are generated from the variation of some particular parameters of the problem (14), see Section 2.2, for more details.

In the next section, the OLU approach is used to generate a suboptimal production plan to the problem (14). The reasons for the choice of the OLU approach are the easiness of the computational implementation, and the possibility of updating information on the state of the system, during the optimization process. Therefore, the OLU approach can provide a production plan that is more accurate than those that no updating any information about the state of the system.

Table 1 summarizes some of the data associated with the parameters and variables of the problem.

Additional information: 1) the standard deviation of the demand is given by and the level of service is set equal to 95%. This level of service means that managers intend to develop a production plan that can satisfy customers for at least 95% of times about delivery term; 2) the index is assumed to be equal to 50%, which means that the production capacity constraint (13) is given by; 3) an ARMA(1,1) model is identified to represent the history of the sales collected monthly for this family of products. From the notation given in (4), the optimal estimated Auto-Regressive and Moving-Average parameters of this model are given by and, respectively; and 4) the HMMS coefficients are given by: C₁ = 69.7; C₂ = 64.3; C₃ = 0.2; C₄ = 5.67; C₅ = 51.2; C₆ = 13.7; C₇ = 0.0825; C₈ = 320; and C₁₂ = 1.134.

Based on the Mean Value Problem (14), the production-planning problem for this example is formulated bellow:

Considering results given at Section 2.3.4; all components and parameters of the problem (15) can be calculated using the data provided previously.

Now, the objective is to find a solution to the problem (15). As discussed in Section 3.2, there are many suboptimal approaches available in the literature. In this paper, the OLU approach is used to solve the problem under study. Besides the computational simplicity, another advantage of this approach is that it allows to incorporate new measures on the inventory and workforce levels over the planning horizon. This last feature is usually known as a rolling horizon scheme, see [13] and [19] , and it is used to update the optimal solution (i.e. the production plan), during the optimization process. Figure 1 illustrates how the OLU approach works to solve the problem (15). Note that for each new period, as soon as new measures are taken from the inventory balance system, the problem (15) is rerun from initial period to end-period. It is worth mentioning that only the optimal policy generated in the period (i.e. and) is effectively applied to the input of the system (see Figure 1). Consequently, the problem (15) must be solved times.

The Figures 2-5 illustrate respectively the optimal trajectories of the inventory, production, workforce and labor-change variables, which are obtained as a result of the use of the OLU approach in the problem (15). Note that they represent the annual production plan, which is desired by the company’s manager. Note also that these optimal trajectories exhibit a particular scenario of the production process, which can be useful for management decision-making.

With respect to trajectories exhibited above, some comments are:

• The trajectory of inventory levels decreased continuously throughout the periods of time, see Figure 2. Note that this characteristic occurs in reason of the use of available information on the production system, during the optimization process. It means that OLU approach allows that the optimization process can be updating with respect to the states of the system, see Figure 1. Note that if for each period the information about the current level of inventory was not incorporated into the process of solution of (15), the tendency of the trajectory of inventory would be to increase continually over the periods of time. The reason of this is that the safety stock, given by function, depends on the variance of the inventory (i.e.,), and so whenever the system operates in an open-loop scheme, the variance of inventory tends to grow with the time. The major implication of this is that the safety stock level (given by) also tends to rise proportionally with time (i.e. periods of the planning horizon). Thus, if the information about the current level of inventory is very weak, then the strategy is to maintain high levels of safety-stock to avoid the risk of stock-

out, see [20] . As a result, it will always be possible to satisfy customers with deliveries on time and still prevent against unexpected events such as broken machines and delays of raw material.

• As illustrated in Figure 3, the optimal production trajectory (solid line) remains relatively stable over the periods, with production levels close to the maximum capacity (i.e.,). Thus, this production policy can answer promptly to the demand fluctuation (dotted line) throughout the periods of planning horizon.

• As a consequence, of the stable behaviour of the production policy, the workforce policy also remains stable for most periods of the planning horizon (see Figure 4). In fact, note that the workforce level fluctuates slightly around the regular workforce level (i.e. workers). Note also that whenever the workforce level overcomes, it is required to adopt one of the two following strategies, that is: subcontracting temporary labor or adopting an overtime policy. Both strategies are responsible for increasing the total production cost but, on the other hand, they allow to keep the production levels close to the maximum capacity of the production process. In short, these strategies improve the competitive performance of the company. Figure 5 ratifies what was said previously, i.e., it shows the fluctuation in the levels of the regular workforce for each period k, where the black bars indicate the level of subcontracting (or overtime), and the white bars indicate the level of the regular workforce.

The results illustrated by Figures 2-5 show one of the possible production scenarios that can be adopted as a target for the short-term planning of the company. It is important to emphasize that the choice of a production scenario that is unique for strategic purposes of the company is not a trivial task. In fact, the main difficulty is related to the need of satisfying tradeoffs such as, for instance: how to minimize inventory levels and, simultaneously, maximize production rates, without introducing temporary workforce or overtime. Another difficulty is due to the lower and upper bounds of inventory and production variables. These constraints reduce the space of feasible solutions to the problem (15) and, consequently, the number of possible scenarios for analysis.

In this section is analyzed the influence of the probabilistic index in the development of scenarios of production. The idea here follows the discussion addressed in Section 2.2, where the index is considered a measure of the customer’s satisfaction. In Section 4.3., the problem (15) was solved with. Using close to 100%, the manager shows a strong commitment of satisfying the customer demand over all periods of the planning horizon. Now, let’s consider that the manager choose equal to 50%. In this case, the manager is assuming the risk of not meeting the demand on time, i.e., he allows backorder. In practice, this type of situation is observed in companies where the demand for innovative products is very high, and the stock in hands becomes obsolete quickly. For this companies, the safety-stock level must be kept as lower as possible, i.e.,

.

In short, using the manager is assuming a more conservative position regarding the administration of inventories, while using, he demonstrates an attitude of always negotiating with customers possible delivery delays. Note that with, it will always be necessary to increase the safety stock levels, particularly in the future periods, to guarantee that the products will be delivered on time to the customers. Raising inventory levels allows the manager to reduce the number of management interventions that are required to adjust the production process, whenever endogenous and exogenous events occur. Examples of such events are unexpected fluctuation of the demand, raw-material delays, machines breakdown, etc. However, it is worth warning that such a managerial approach implies increasing the total production cost.

Figure 6 and Figure 7 shows the inventory and production trajectories for both (dashed line) and 95% (solid lines). From Figure 6, note that the inventory levels related to are slightly superior to the ones related to. This feature shows that the use of probabilistic indexes near to 100% increase the inventory levels as previously discussed. Indeed, the safety stock level increases over the periods of the planning horizon to guarantee that future demands will be met. Therefore, the total production cost for the policy with is more expensive (around 3%) than the one provided by. In fact, looking carefully over these production trajectories, it is possible to verify that the production policy with, remains 70% of times operating exactly in its maximum capacity, while the production policy, with, operates only 25% of the time in a maximum capacity. In the case of, this characteristic means that the whole production capacity is geared to meet the demand.

Finally, it is interesting to mention that the development of production scenarios, based on safety stock levels to satisfy customer demand, has strong interest for those companies that operate in markets of commodities, i.e., companies whose products are made to stocks.