betaFeedback

WorksMathematics

Two-Stage Staffing with Random Arrival Rate: A Simulation Optimization Approach

Abstract

In service systems, the arrival rate in the same time period and day of the week exhibits randomness, yet can be more accurately estimated or forecast right before that period, which motivates a two-stage staffing scheme with advance base staffing and pre-shift surge staffing. Existing approaches usually impose strong assumptions on the arrival process and service systems' operational logic and other further approximations (i.e., stylized models) to derive explicit staffing policies. This paper proposes a simulation optimization approach (SOA) for more general arrival processes and service systems by effectively addressing some intertwined challenges. By recasting the original formulation into computationally tractable structures and designing an efficient coordinate-search algorithm, it constructs a two-stage staffing lookup table offline, which allows managers to make real-time surge staffing decisions without any further simulation. Comprehensive numerical experiments demonstrate that the proposed SOA has significantly better staffing performance and substantially broader applicability compared to existing approaches. Even under the stylized models, the cost savings delivered by our SOA compared with existing approaches range from the low tens to several tens of percent, depending on different problem parameters; the advantage further amplifies when the actual problem deviates from the stylized models. In a nurse-staffing case study grounded in real-world data, the proposed SOA delivers approximately 8% to 13% cost savings over a state-of-the-art benchmark.

Introduction

In many service systems, staffing decisions play a central role in matching service capacity to demand. For example, in emergency departments (EDs), inadequate nurse staffing is associated with longer patient lengths of stay and higher leaving without being seen (LWBS) rates [4]. Similarly, in call centers, insufficient agent staffing can increase customer waiting time and abandonment rates [6]. Conversely, excess staffing leaves capacity underutilized, with high labor costs accounting for a large proportion of operational expenses. Therefore, balancing staffing levels (and cost) against quality of service (QoS) is a key issue in service systems.

However, making effective staffing decisions is challenging, largely because the service demand is often uncertain and varies over time. For example, although the arrivals in healthcare systems or call centers may exhibit certain within-day and within-week patterns, accurately predicting the number of arrivals or timing of arrivals in a given time period is very challenging [7]. Empirical studies show that the number of arrivals in the same time period and day of the week across multiple weeks often exhibits overdispersion, i.e., the variance is significantly greater than the mean, which suggests that a standard non-homogeneous Poisson process is often not capable of characterizing the real arrival process [13]. Many studies, e.g., staffing in healthcare systems (Hu et al., 2025) and call centers [18] (Koçağa et al. 2015), tried to model the arrival process as a doubly stochastic Poisson process (DSPP), where the arrival rate of the Poisson process during some time period is a random variable.

In practice, staffing decisions in service systems often need to be made weeks in advance. But if the arrival rate during some time period itself is random, the staffing level planned based on the mean (i.e., expectation) of the arrival rate in advance must mismatch the actual arrival rate realized when that day comes [3] (Koçağa et al. 2015). On the other hand, research indicates that short-term forecasting of arrival rates can incorporate a wealth of real-time information, thereby achieving superior predictive accuracy compared to long-term forecasting [19] (Hu et al. 2023). This naturally motivates a two-stage staffing scheme, where the staffing plan made in advance can be adjusted once the arrival rate is more accurately estimated or forecast, via overtime or temporary staff, albeit at higher unit cost. For example, Koçağa et al. (2015) consider call centers where managers first determine in-house staffing and then, once the arrival rate is known, decide whether to outsource incoming calls based on system congestion. Kim and Mehrotra [12] use two-stage stochastic programming to optimize both nurse staffing and scheduling, with initial plans adjusted in response to updated demand forecasts. Ta et al. (2021) study staffing in call centers, where staffing level is first determined according to the distribution of arrival rate and adjusted in the second stage with updated forecasts, to minimize the staff cost while maintaining certain QoS.

A more recent and notable work is Hu et al. (2025), which considers the two-stage nurse staffing in EDs. Patient arrivals are assumed to follow a DSPP, and the ED is assumed to operate as an M/M/k+MM/M/k+M queue given the realized arrival rate and specific QoS is considered. Under this setting, they derive an explicit two-stage quality-and-efficiency driven (QED) staffing policy for the single-shift case based on some analytical approximations, with base staffing planned in advance and surge staffing determined once the arrival rate of that shift is known. They prove that the two-stage QED policy is near-optimal in an asymptotic sense for the single-shift case, and suggest some heuristic adjustments for the multi-shift case. Numerical experiments including a case study with real data demonstrate that the policy has satisfactory performance. Nevertheless, several limitations remain. First, the assumptions of the stylized models for the arrival process and service system limit its application in a more realistic or broader setting. Second, the asymptotic optimality property does not necessarily ensure optimality in reality, when the service system operates in a transient state due to the time-varying arrivals and resources. Third, the adjustments for the multi-shift case require coefficient calibration, making its performance potentially problem instance-specific.

In this paper, we propose a simulation optimization approach (SOA) framework for general two-stage staffing problems with random arrival rates, including, but not limited to, the specific nurse staffing in EDs considered in Hu et al. (2025). It is motivated by the wide application of simulation optimization in operations management [2]. For staffing problems, Feldman et al. [5] propose a simulation-based iterative staffing algorithm for time-varying queues to achieve time-stable performance. In healthcare, simulation optimization has been widely used for resource planning, including ED resource allocation [1], long-term care capacity planning [21], and ED physician staffing and scheduling [17]; see Yousefi et al. [20] and Wang and Demeulemeester [16] for a review. These studies show the value of simulation optimization in evaluating and improving staffing and resource allocation decisions, when complicated system dynamics make analytical solutions difficult to obtain. To the best of our knowledge, however, the literature remains silent on directly applying simulation optimization to the two-stage staffing problem with random arrival rates.

The most immediate benefit of the SOA is that it allows general arrival processes with random rates and service systems that operate under arbitrary yet well-defined operational logic, not limited to the stylized models. However, applying simulation optimization to such a two-stage staffing problem is far from trivial and inherently accompanied by critical challenges. First, as will be elaborated in the sequel, the original formulation of the two-stage staffing problem for the general multi-shift setting constitutes a highly complex dynamic program. Even leaving aside the time required to estimate the objective function via simulation, a direct solution to this dynamic programming problem is practically intractable. Second, the simulation time remains a non-negligible factor, despite utilizing robust computational resources and efficient techniques. Third, even if solving the simulation optimization is computationally viable within a reasonable time, the second-stage surge staffing decision must be made rapidly after the arrival rate becomes known. Such a strict (near-)real-time requirement is inherently beyond the reach of instantaneous simulation optimization. These intertwined challenges are effectively addressed in this paper.

The main contributions of this paper can be highlighted as follows. To the best of our knowledge, this study represents the first attempt to leverage simulation optimization directly for the general two-stage staffing problem under random arrival rates, accomplished through our proposed SOA framework. To deal with the computational intractability of the original dynamic program and the real-time requirement for surge staffing decisions, we recast the multi-shift problem via tractable approximation, solve the resulting simulation optimization efficiently through tailored coordinate-search algorithms, and shift the entire computational burden offline. Specifically, offline-generated arrival-rate paths are utilized to pre-compute the surge staffing decisions across arrival-rate intervals, yielding a two-stage staffing lookup table. Once the true arrival rate is revealed prior to a shift, managers can instantaneously retrieve the surge staffing decision in real time without any further simulation, just like using closed-form formulas. Comprehensive numerical experiments demonstrate that the proposed SOA has significantly better staffing performance and substantially broader applicability compared to existing approaches. Moreover, our study reveals when the complexity of two-stage staffing will offer meaningful gains over the simple one-stage schedule.

The rest of the paper is organized as follows. In Section 2 we present the original formulation of the two-stage staffing problem in detail and discuss the main difficulty for solving it. In Section 3, the SOA for the single-shift case and general multi-shift case is presented, respectively, followed by the comparison with existing policies. Section 4 includes comprehensive numerical experiments, demonstrating the performance and superiority of the SOA, and Section 5 briefly concludes the paper. Some additional numerical results are collected in the Appendix.

Problem Formulation

We consider a general two-stage staffing problem in a service system. Suppose there are totally II consecutive shifts, and the initial state of the first shift is given as π0\pi_0. For a given shift ii, i=1,…,Ii = 1,\ldots,I, the demand arrives following a certain counting process, where λi\lambda_i is a constant arrival rate during shift ii. In principle, the service system can operate under arbitrary yet well-defined operational logic. Usually, one may consider a multi-server queuing system, possibly with some specific features, including queue discipline (first-come-first-served, shortest-processing-time-first, etc.), queue behavior (balking, reneging, etc.), and so on. However, when deciding the base staffing level (i.e., number of servers) b:=(b1,…,bI)∈NI\boldsymbol{b} := (b_1,\ldots,b_I) \in\mathbb{N}^I some time in advance, the actual arrival rates are unknown (or unrealized) yet.

The arrival rates λi\lambda_i, i=1,…,Ii = 1,\ldots,I, are modeled as random variables Λi\Lambda_i with marginal distribution FiF_i, i=1,…,Ii = 1,\ldots,I. We assume that their joint distribution FF is known and allow the arrival rates to be correlated across shifts. Following Hu et al. (2025), this paper simplifies the environment by assuming that the actual arrival rates are perfectly revealed prior to each shift. We recognize, however, that in practice they must be forecast with real-time information in some way, and thus are inevitably subject to errors, despite being substantially more accurate than simple expected values (Hu et al., 2023). Incorporating such forecasting errors into the model is a valuable topic for future research, but is beyond the scope of this study. For ease of exposition, in the sequel we will simply state that λi\lambda_i is known or observed prior to the beginning of shift ii.

Suppose the base staffing level b\boldsymbol{b} has been decided. Prior to the beginning of shift 1, the actual arrival rate λ1\lambda_1 is known and the surge staffing level is decided to be s1∈Ns_1 \in\mathbb{N}, where one surge staff is the same as one base staff in terms of service rate and quality, but costs more. When shift 1 begins, the total staffing level is k1=b1+s1k_1 = b_1 + s_1. It reflects two operational issues in practice. One is that it is difficult to downsize scheduled base staff at short notice, and the other is that it costs more to add staff at short notice. Then prior to the beginning of shift 2, the actual arrival rate λ2\lambda_2 is known, together with the initial state of shift 2 (i.e., the terminating state of shift 1) that is denoted as π1\pi_1, and surge staffing level s2∈Ns_2 \in\mathbb{N} needs to be determined. It follows likewise for i=3,…,Ii = 3,\ldots,I. summarize, the base staffing level b\mathbf{b} is determined and fixed in advance with knowledge of the joint distribution FF, the surge staffing level s:=(s1,…,sI)∈NI\mathbf{s} := (s_1,\ldots,s_I) \in\mathbb{N}^I is decided sequentially, where each sis_i is determined upon knowing λi\lambda_i and πi−1\pi_{i-1} prior to the beginning of shift ii, and the total staffing level ki=bi+sik_i = b_i + s_i is actually set in shift ii, for i=1,…,Ii = 1,\ldots,I.

The objective is to minimize the sum of the cost of staffing and the expected penalty related to the QoS of the service system. Denote fi(ki,λi,πi−1)f_i(k_i,\lambda_i,\pi_{i-1}) as the expected penalty related to the QoS of the service system with initial state πi−1\pi_{i-1} when the actual staffing level is kik_i and the arrival rate during the shift is λi\lambda_i. Here the expectation is taken with respect to the randomness in the queuing system in shift ii. And denote ci(b)c_i^{(b)} and ci(s)c_i^{(s)} as the unit cost of base and surge staff, respectively, for i=1,…,Ii = 1,\ldots,I. Note that fi(⋅)≥0f_i(\cdot) \ge0 and ci(s)>ci(b)>0c_i^{(s)} > c_i^{(b)} > 0.

Simplified Formulation for Single Shift

Let us first consider the simplified case where I=1I = 1, or treat the service system as if the II (I>1I > 1) shifts were independent and the initial state of each shift were given. In the latter case, the optimization problem just degenerates to II isolated single-shift problems with a known initial state. In either case, we only need to focus on one shift, say ii. Upon knowing λi\lambda_i, the second-stage optimization problem is defined as

z(bi,λi,πi−1):=min⁡si∈N{ci(s)si+fi(bi+si,λi,πi−1)},(1)z(b_i,\lambda_i,\pi_{i-1}) := \min_{s_i \in\mathbb{N}} \left\{c_i^{(s)}s_i + f_i(b_i+s_i,\lambda_i,\pi_{i-1})\right\}, \tag*{(1)}

and the first-stage optimization problem is defined as

bi∗:=arg⁡min⁡bi∈N{ci(b)bi+EΛi[z(bi,Λi,πi−1)]},(2)b_i^* := \arg\min_{b_i \in\mathbb{N}} \left\{c_i^{(b)}b_i + \mathbb{E}_{\Lambda_i}\left[z(b_i,\Lambda_i,\pi_{i-1})\right]\right\}, \tag*{(2)}

where the expectation is taken with respect to the random arrival rate Λi∼Fi\Lambda_i \sim F_i. Given the optimal base staffing level bi∗b_i^*, the optimal surge staffing level upon knowing λi\lambda_i is

si∗:=arg⁡min⁡si∈N{ci(s)si+fi(bi∗+si,λi,πi−1)}.(3)s_i^* := \arg\min_{s_i \in\mathbb{N}} \left\{c_i^{(s)}s_i + f_i(b_i^*+s_i,\lambda_i,\pi_{i-1})\right\}. \tag*{(3)}

General Formulation for Multiple Shifts

Unfortunately, II shifts are usually dependent when I>1I > 1, as the unfinished customers in one shift usually carry over to the next shift in practical service systems. That is, the initial state of shift ii, πi−1\pi_{i-1}, is not given in advance, but a consequence of the arrival rates and staffing decisions in all previous shifts (subject to randomness in the queuing system). Let Πi−1\Pi_{i-1} denote the random initial state of shift ii. Suppose the base staffing level b∈NI\boldsymbol{b} \in\mathbb{N}^{I} has been decided. Let YiY_i denote the available information besides b\boldsymbol{b} when deciding the surge staffing level for shift ii, and let SiS_i denote the corresponding surge staffing decision (which is also a random variable determined by YiY_i and b\boldsymbol{b}). Then we can see that

Y1:=(π0,Λ1),Yi+1:=(Yi,Si,Πi,Λi+1),i=1,…,I−1.\begin{aligned} Y_1 &:= (\pi_0, \Lambda_1),\\ Y_{i+1} &:= (Y_i, S_i, \Pi_i, \Lambda_{i+1}), \quad i = 1, \ldots, I - 1. \end{aligned}

Note that Πi\Pi_i is actually determined by Πi−1\Pi_{i-1}, Λi\Lambda_i, and bi+Sib_i + S_i, together with the randomness in the queuing system in shift ii.

Suppose YiY_i is realized as yiy_i, where Πi−1\Pi_{i-1} is realized as πi−1\pi_{i-1} and Λi\Lambda_i is realized as λi\lambda_i. Let Vi(yi;b)V_i(y_i; \boldsymbol{b}) denote the expected remaining cumulative cost from shift ii until shift II if the surge staffing levels for all (I−i+1)(I-i+1) shifts are sequentially and optimally decided. Let VI+1(⋅;b)=0V_{I+1}(\cdot; \boldsymbol{b}) = 0, and then for i=1,…,Ii = 1, \ldots, I, we have the recursive relationship:

Vi(yi;b)=min⁡si∈N{ci(s)si+fi(bi+si,λi,πi−1)+EΠi,Λi+1[Vi+1(Yi+1;b)∣Yi=yi,Si=si]}V_i(y_i; \boldsymbol{b}) = \min_{s_i \in\mathbb{N}} \left\{c_i^{(s)}s_i + f_i(b_i+s_i,\lambda_i,\pi_{i-1}) + \mathbb{E}_{\Pi_i,\Lambda_{i+1}}\left[V_{i+1}(Y_{i+1};\boldsymbol{b}) \mid Y_i=y_i,S_i=s_i\right]\right\}

where the expectation is taken with respect to the random initial state Πi\Pi_i and random arrival rate Λi+1\Lambda_{i+1} in shift (i+1)(i+1). Then the first-stage optimization problem for deciding the optimal base staffing level b\boldsymbol{b} is

b∗:=arg⁡min⁡b∈NI{∑i=1Ici(b)bi+EΛ1[V1(Y1;b)]}(4)\boldsymbol{b}^{*} := \arg\min_{\boldsymbol{b}\in\mathbb{N}^{I}}\left\{\sum_{i=1}^{I}c_i^{(b)}b_i+\mathbb{E}_{\Lambda_1}\left[V_1(Y_1;\boldsymbol{b})\right]\right\} \tag*{(4)}

where the expectation is taken with respect to the random arrival rate Λ1\Lambda_1 in shift 1. Note that the optimization of surge staffing is still performed sequentially, as the actual arrival rates are revealed sequentially. Upon knowing yiy_i prior to the beginning of shift ii, i=1,…,Ii = 1, \ldots, I, the optimal surge staffing level for shift ii is

si∗:=arg⁡min⁡si∈N{ci(s)si+fi(bi∗+si,λi,πi−1)+EΠi,Ai+1[Vi+1(Yi+1;b∗)∣Yi=yi,Si=si]}(5)s_i^* := \underset{s_i \in\mathbb{N}}{\arg\min} \left\{ c_i^{(s)}s_i + f_i(b_i^* + s_i, \lambda_i, \pi_{i-1}) + \mathbb{E}_{\Pi_i,A_{i+1}}\left[V_{i+1}(Y_{i+1};\mathbf{b}^*) \mid Y_i=y_i,S_i=s_i\right] \right\} \tag*{(5)}

The Main Difficulty

The above optimization problems are mathematically accurate, but it is almost impossible to directly solve them. First, consider the simplified formulation for the single-shift case. The main difficulty lies in the intractability of fi(ki,λi,πi−1)f_i(k_i,\lambda_i,\pi_{i-1}). Over a finite shift horizon (e.g., 8 or 12 hours), transient dynamics and the initial state πi−1\pi_{i-1} play an important role, so the system cannot be treated as operating in steady state. Thus the expected penalty related to the QoS of the service system is difficult to derive, even if we assume some stylized queuing models with well-known steady-state distributions. And in fact some specific queue features, such as queue discipline and queue behavior, need to be considered in the service system in order to be realistic. It makes the mathematical derivation of fi(⋅)f_i(\cdot) even less likely. Approximation and simulation-based estimation seem to be the only two possible methods.

When it comes to the general formulation for the multi-shift case, this formulation significantly magnifies the difficulty. First, the system state will transit among consecutive shifts, which means the distribution of system states in different shifts should be analyzed jointly. An exact evaluation is extremely difficult; even a simulation-based estimation is highly challenging, to say nothing of theoretical analysis or approximation. Moreover, the defined expected remaining cumulative cost Vi(yi;b)V_i(y_i;\mathbf{b}) involves both challenges of expectation estimation and iterative optimization. Solving such a dynamic programming problem directly is practically intractable for moderate values of II, even if we can efficiently simulate the service system.

Hu et al. (2025) take an analytical approximation approach to two-stage staffing. Specifically, they impose stylized models for the arrival process and service system with specific QoS considered (which will be elaborated in detail later). For the single-shift case, by further assuming the system operates in a steady state, they first derive a two-stage Newsvendor rule using a stochastic-fluid approximation and then refine it using a diffusion approximation to obtain the two-stage QED staffing rule. For the general multi-shift case, some heuristic adjustments to the previous staffing rule are suggested to compensate for the transient-shift effect. In this paper, we take a simulation optimization approach for much more general settings. As previously noted and elaborated further below, applying simulation to this problem—particularly in the multi-shift case—is not straightforward, requiring an appropriate reformulation of the original problem.

Simulation Optimization Approach for Two-Stage Staffing

In this section, we first present the SOA for the single-shift case, then extend it to the full version for the more general multi-shift case. We will also compare the SOA with existing approaches in terms of both problem setting and staffing policy.

SOA for Single Shift

Recall that fi(ki,λi,πi−1)f_i(k_i,\lambda_i,\pi_{i-1}) is the expected penalty related to the QoS of the service system with initial state πi−1\pi_{i-1} when the actual staffing level is kik_i and the arrival rate during the shift is λi\lambda_i. Since the service system will usually not enter a steady state during one shift, fi(⋅)f_i(\cdot) should be estimated by running terminating simulation, where the time duration is the shift length. Note that πi−1\pi_{i-1} is given and fixed in the single-shift case.

To solve the optimization problems (2) and (3), we adopt the classical two-layer nested simulation. In the outer layer, MM values of Λi\Lambda_i are sampled from the distribution of Λi\Lambda_i, FiF_i, which are denoted as λim\lambda_{im}, m=1,…,Mm=1,\ldots,M. In the inner layer, fi(ki,λi,πi−1)f_i(k_i,\lambda_i,\pi_{i-1}) is estimated for given kik_i and λi\lambda_i. Let fir(ki,λi,πi−1)f_{ir}(k_i,\lambda_i,\pi_{i-1}) denote the calculated penalty in the rr-th terminating simulation run, and totally RR independent simulation replications are conducted. Then fi(ki,λi,πi−1)f_i(k_i,\lambda_i,\pi_{i-1}) is estimated by

fˉiR(ki,λi,πi−1):=1R∑r=1Rfir(ki,λi,πi−1).\bar{f}_{iR}(k_i,\lambda_i,\pi_{i-1}) := \frac{1}{R}\sum_{r=1}^{R} f_{ir}(k_i,\lambda_i,\pi_{i-1}).

To more efficiently compare fˉiR(ki,λi,πi−1)\bar{f}_{iR}(k_i,\lambda_i,\pi_{i-1}) for different values of kik_i, one can use common random numbers (CRN) across different values of kik_i under the same λi\lambda_i. Specifically, the simulations with different values of kik_i share the same set of exogenous random inputs, including random arrivals, as well as their service times and other personal characteristics.

Given the first-stage decision bib_i, for each outer-layer realization λim\lambda_{im}, the second-stage optimization problem becomes

zˉR(bi,λim,πi−1):=min⁡si∈N{ci(s)si+fˉiR(bi+si,λim,πi−1)}.(6)\bar{z}_{R}(b_i,\lambda_{im},\pi_{i-1}) := \min_{s_i\in\mathbb{N}} \left\{c_i^{(s)}s_i+\bar{f}_{iR}(b_i+s_i,\lambda_{im},\pi_{i-1})\right\}. \tag*{(6)}

So, the first-stage optimization problem for deciding the optimal base staffing level bib_i in SOA is

bˉi∗:=arg⁡min⁡bi∈N{ci(b)bi+1M∑m=1MzˉR(bi,λim,πi−1)}.(7)\bar{b}_i^* := \arg\min_{b_i\in\mathbb{N}} \left\{c_i^{(b)}b_i+\frac{1}{M}\sum_{m=1}^{M}\bar{z}_{R}(b_i,\lambda_{im},\pi_{i-1})\right\}. \tag*{(7)}

The optimal surge staffing level upon knowing λi\lambda_i in SOA is

sˉi∗:=arg⁡min⁡si∈N{ci(s)si+fˉiR(bˉi∗+si,λi,πi−1)}.(8)\bar{s}_i^* := \arg\min_{s_i\in\mathbb{N}} \left\{c_i^{(s)}s_i+\bar{f}_{iR}(\bar{b}_i^*+s_i,\lambda_i,\pi_{i-1})\right\}. \tag*{(8)}

Since both Problems (6)-(7) are one-dimensional integer simulation optimization problems, we simply employ a hill-climbing search strategy. That is, to solve min⁡xg(x)\min_x g(x) with currently found solution xx, iteratively move to the neighbor (x−1x-1 or x+1x+1) with the smallest estimated function value, if it is lower than the estimated value of g(x)g(x); and terminate otherwise. For each candidate base staffing level bib_i, we use an inner hill-climbing search to find a local solution in the second-stage Problem (6) under each of the MM realizations of Λi\Lambda_i, where each fˉiR\bar{f}_{iR} is computed from RR simulation replications of the service system. With each estimated zˉR(bi,λim,πi−1)\bar{z}_{R}(b_i,\lambda_{im},\pi_{i-1}), we can estimate the objective function in the first stage Problem (7) for each bib_i. Based on this, an outer hill-climbing search is used to solve bˉi∗\bar{b}_i^*. In the implementation, to avoid redundant simulation, we cache the simulation estimates fˉiR(ki,λi,πi−1)\bar{f}_{iR}(k_i,\lambda_i,\pi_{i-1}) by (ki,λi)(k_i,\lambda_i). CRN is used when evaluating fˉiR(ki,λi,πi−1)\bar{f}_{iR}(k_i,\lambda_i,\pi_{i-1}) for different values of kik_i under the same λi\lambda_i. Moreover, for each bib_i, the inner searches across the MM realizations of Λi\Lambda_i can be performed in parallel.

Solving sˉi∗\bar{s}_i^* in (8) is just similar to solving Problem (6) once with bi=bˉi∗b_i=\bar{b}_i^* and λim=λi\lambda_{im}=\lambda_i. So the computation delay may not significantly impede the near-real-time decision making upon knowing λi\lambda_i prior to the beginning of shift ii. If strict real-time decision-making is required, one can adopt the so-called offline simulation online application idea [9]. That is to say, pre-compute sˉi∗\bar{s}_i^* for several values of λi\lambda_i before observing it (i.e., solve (8) for different values of λi\lambda_i), and report the most applicable sˉi∗\bar{s}_i^* upon observing λi\lambda_i. We will illustrate this issue in detail and adopt a simple method in the multi-shift case, where the computation delay indeed constitutes a non-trivial barrier for (near-)real-time decision making.

SOA for Multiple Shifts

Note that directly solving the optimization problems (4) and (5) is practically intractable, because the computation burden of both simulation and optimization will be easily prohibitive for moderate values of II. To tackle this challenge, an approximation of the original optimization is considered.

The Approximated Model

To calculate the optimal base staffing level b\mathbf{b}, consider a more tractable scenario, where all actual arrival rates are revealed simultaneously (instead of sequentially). For given actual staffing level k:=(k1,…,kI)\mathbf{k}:=(k_1,\ldots,k_I) and arrival rate λ:=(λ1,…,λI)\boldsymbol{\lambda}:=(\lambda_1,\ldots,\lambda_I), define

g(k,λ,π0):=EΠ1,…,ΠI−1[∑i=1Ifi(ki,λi,Πi−1)∣Π0=π0](9)g(\mathbf{k},\boldsymbol{\lambda},\pi_0):=\mathbb{E}_{\Pi_1,\ldots,\Pi_{I-1}}\left[\left.\sum_{i=1}^{I}f_i(k_i,\lambda_i,\Pi_{i-1})\right|\Pi_0=\pi_0\right] \tag*{(9)}

which means the entire cumulative expected penalty related to the QoS of the service system. Note that the expectation in (9) is taken with respect to (Π1,…,ΠI−1)(\Pi_1,\ldots,\Pi_{I-1}), since they are random for given k\mathbf{k} and λ\boldsymbol{\lambda} due to the randomness in the queuing system. Then, the optimization problem for deciding the optimal base staffing level b\mathbf{b} becomes

b~∗:=arg⁡min⁡b∈NI{∑i=1Ici(b)bi+EΛ[min⁡s∈NI{∑i=1Ici(s)si+g(b+s,Λ,π0)}]}(10)\widetilde{\mathbf{b}}^{*}:=\arg\min_{\mathbf{b}\in\mathbb{N}^{I}}\left\{\sum_{i=1}^{I}c_i^{(b)}b_i+\mathbb{E}_{\boldsymbol{\Lambda}}\left[\min_{\mathbf{s}\in\mathbb{N}^{I}}\left\{\sum_{i=1}^{I}c_i^{(s)}s_i+g(\mathbf{b}+\mathbf{s},\boldsymbol{\Lambda},\pi_0)\right\}\right]\right\} \tag*{(10)}

where the expectation is taken with respect to the random arrival rate vector Λ:=(Λ1,…,ΛI)\boldsymbol{\Lambda}:=(\Lambda_1,\ldots,\Lambda_I).

Although the base staffing level is decided by treating the arrival rates as if they were revealed simultaneously, the surge staffing level still needs to be decided sequentially. Prior to the beginning of shift ii, λi\lambda_i and πi−1\pi_{i-1} will be observed. Strictly speaking, the optimal sis_i must depend on both λi\lambda_i and πi−1\pi_{i-1}. But πi−1\pi_{i-1} could be a set with tremendous variables (which together define the state of a system), in the simulation viewpoint explicit dependence on πi−1\pi_{i-1} means we have to discard all the previous simulated sample paths, and restart from πi−1\pi_{i-1} to the final shift. Such computation delay will make the (near-)real-time decision making impossible. To reduce the computation burden, we use an approximation that avoids re-optimizing surge staffing for the observed initial state. More specifically, let

s(b~∗,λ,π0):=arg⁡min⁡s∈NI{∑i=1Ici(s)si+g(b~∗+s,λ,π0)}.s(\widetilde{b}^{*},\boldsymbol{\lambda},\pi_{0}) := \arg\min_{s\in\mathbb{N}^{I}} \left\{\sum_{i=1}^{I}c_i^{(s)}s_i+g(\widetilde{b}^{*}+s,\boldsymbol{\lambda},\pi_{0})\right\}.

and let si(b~∗,λ,π0)s_i(\widetilde{b}^{*},\boldsymbol{\lambda},\pi_{0}) denote the ii-th element of s(b~∗,λ,π0)s(\widetilde{b}^{*},\boldsymbol{\lambda},\pi_{0}). Prior to the beginning of shift ii, we observe λi\lambda_i and decide the surge staffing level as

s~i∗:=round⁡(E[si(b~∗,Λ,π0)∣Λi=λi]).(11)\widetilde{s}_i^{*}:=\operatorname{round}\left(\mathbb{E}\left[s_i(\widetilde{b}^{*},\boldsymbol{\Lambda},\pi_{0})\mid\Lambda_i=\lambda_i\right]\right). \tag*{(11)}

Simulation-Based Solution

Similar to the single-shift case, (10) needs to be solved using two-layer nested simulation. In the outer layer, MM values of Λ\boldsymbol{\Lambda} are sampled from the joint distribution of Λ\boldsymbol{\Lambda}, which are denoted as λm\boldsymbol{\lambda}_m, m=1,…,Mm=1,\ldots,M. In the inner layer, g(k,λ,π0)g(\boldsymbol{k},\boldsymbol{\lambda},\pi_0) is estimated for given k\boldsymbol{k} and λ\boldsymbol{\lambda} with RR independent simulation replications. A convenient way to estimate g(k,λ,π0)g(\boldsymbol{k},\boldsymbol{\lambda},\pi_0) is to generate the entire sample path from shift 1 to II with given k\boldsymbol{k} and λ\boldsymbol{\lambda}, for RR times. Each sample path tracks the complete trajectory of the service system state over time, based on which a random sample of g(k,λ,π0)g(\boldsymbol{k},\boldsymbol{\lambda},\pi_0) can be computed and denoted as Gr(k,λ,π0)G_r(\boldsymbol{k},\boldsymbol{\lambda},\pi_0). Then g(k,λ,π0)g(\boldsymbol{k},\boldsymbol{\lambda},\pi_0) is estimated by

g‾R(k,λ,π0):=1R∑r=1RGr(k,λ,π0).\overline{g}_R(\boldsymbol{k},\boldsymbol{\lambda},\pi_0):=\frac{1}{R}\sum_{r=1}^{R}G_r(\boldsymbol{k},\boldsymbol{\lambda},\pi_0).

Thus, the optimization problem for deciding the optimal base staffing level b\boldsymbol{b} in SOA is

b~∗:=arg⁡min⁡b∈NI{∑i=1Ici(b)bi+1M∑m=1M[min⁡s∈NI{∑i=1Ici(s)si+g‾R(b+s,λm,π0)}]}.(12)\widetilde{b}^{*}:=\arg\min_{\boldsymbol{b}\in\mathbb{N}^{I}}\left\{\sum_{i=1}^{I}c_i^{(b)}b_i+\frac{1}{M}\sum_{m=1}^{M}\left[\min_{s\in\mathbb{N}^{I}}\left\{\sum_{i=1}^{I}c_i^{(s)}s_i+\overline{g}_R(\boldsymbol{b}+s,\boldsymbol{\lambda}_m,\pi_0)\right\}\right]\right\}. \tag*{(12)}

To solve b~∗\widetilde{b}^{*} in (12), we adopt the coordinate-search algorithm proposed by Hong [8], which is a coordinate-wise hill-climbing search algorithm. Define

hm(b,s):=∑i=1Ici(s)si+g‾R(b+s,λm,π0),h_m(\boldsymbol{b},s):=\sum_{i=1}^{I}c_i^{(s)}s_i+\overline{g}_R(\boldsymbol{b}+s,\boldsymbol{\lambda}_m,\pi_0),

and sm(b):=arg⁡min⁡s∈NIhm(b,s)s_m(\boldsymbol{b}):=\arg\min_{s\in\mathbb{N}^{I}}h_m(\boldsymbol{b},s). The entire procedure is summarized in Algorithm 1. The implementation strategies to enhance computational efficiency are analogous to those used in the singleshift case, including caching gˉR(k,λ,π0)\bar{g}_R(k,\boldsymbol{\lambda},\pi_0) by (k,λ)(k,\boldsymbol{\lambda}), utilizing CRN when evaluating gˉR(k,λ,π0)\bar{g}_R(k,\boldsymbol{\lambda},\pi_0) for different values of kk under the same λ\boldsymbol{\lambda}, and parallelizing for λm\boldsymbol{\lambda}_m, m=1,…,Mm=1,\ldots,M.

Strictly speaking, to estimate s~i∗\tilde{s}_i^* in (11), one needs to select λm∗\boldsymbol{\lambda}_m^*s from the MM generated samples of Λ\boldsymbol{\Lambda} such that their ii-th element exactly equals to λi\lambda_i, and then calculate the average of si(b^∗,λ,π0)s_i(\hat{\boldsymbol{b}}^*,\boldsymbol{\lambda},\pi_0) over these λm∗\boldsymbol{\lambda}_m^*s. However, due to the continuity of Λ\boldsymbol{\Lambda}, the probability of existence of such λm∗\boldsymbol{\lambda}_m^*s is zero. Then one has to re-generate samples of Λ\boldsymbol{\Lambda} conditional on Λi=λi\Lambda_i=\lambda_i and compute si(b^∗,λ,π0)s_i(\hat{\boldsymbol{b}}^*,\boldsymbol{\lambda},\pi_0) for each of them. This computation can only start upon observing λi\lambda_i, whose time delay will violate the requirement of (near-)real-time decision making.

We propose to further simplify the computation by utilizing the offline-generated samples and computed results. Let b^\hat{\boldsymbol{b}} be the solved base staffing and sm(b^)s_m(\hat{\boldsymbol{b}}) the associated surge staffing solutions if actual arrival rates are revealed simultaneously as λm\boldsymbol{\lambda}_m, m=1,…,Mm=1,\ldots,M, which are returned by Algorithm 1. Denote λmi\lambda_{mi} as the ii-th element of λm\boldsymbol{\lambda}_m. Divide the possible range of arrival rate (i.e., (0,∞)(0,\infty)) into NN disjoint intervals based on the quantiles of {λ1i,…,λMi}\{\lambda_{1i},\ldots,\lambda_{Mi}\}. More specifically, order

Algorithm 1 Coordinate-Search Simulation Optimization Algorithm for Base Staffing

Require: Arrival rate realizations {λm}m=1M\{\boldsymbol{\lambda}_m\}_{m=1}^{M}, initial state π0\pi_0; outer search round KoutK_{\mathrm{out}}, inner search round KinK_{\mathrm{in}}; initial solution b(0)\boldsymbol{b}^{(0)}.

Ensure: Base staffing solution b^\hat{\boldsymbol{b}}, corresponding joint surge solutions {sm(b^)}m=1M\{s_m(\hat{\boldsymbol{b}})\}_{m=1}^{M}.

Main Function in Outer Layer

  1. b←b(0)\boldsymbol{b}\leftarrow\boldsymbol{b}^{(0)}

  2. for kout=1,…,Koutk_{\mathrm{out}}=1,\ldots,K_{\mathrm{out}} do

  3.     for i=1,…,Ii=1,\ldots,I do

  4.         Update bib_i while fixing other elements in b\boldsymbol{b} by one-dimensional hill-climbing search, via calling Evaluate(b)(\boldsymbol{b}) to estimate the objective function for each b\boldsymbol{b}.

  5.     end for

  6. end for

  7. b^←b\hat{\boldsymbol{b}}\leftarrow\boldsymbol{b}.

  8. return b^\hat{\boldsymbol{b}} and {sm(b^)}m=1M\{s_m(\hat{\boldsymbol{b}})\}_{m=1}^{M}.

Objective Function Estimation in Inner Layer

  1. function Evaluate(b)(\boldsymbol{b})

  2.     for m=1,…,Mm=1,\ldots,M in parallel do

  3.         Initialize sm\boldsymbol{s}_m.

  4.         for kin=1,…,Kink_{\mathrm{in}}=1,\ldots,K_{\mathrm{in}} do

  5.             for j=1,…,Ij=1,\ldots,I do

  6.                 Update smjs_{mj} while fixing other elements in sm\boldsymbol{s}_m by one-dimensional hill-climbing search on hm(b,sm)h_m(\boldsymbol{b},\boldsymbol{s}_m).

  7.             end for

  8.         end for

  9.     end for

  10. return ∑i=1Ici(b)bi+1M∑m=1Mhm(b,sm)\sum_{i=1}^{I}c_i^{(b)}b_i+\frac{1}{M}\sum_{m=1}^{M}h_m(\boldsymbol{b},\boldsymbol{s}_m), and associated solutions {sm}m=1M\{\boldsymbol{s}_m\}_{m=1}^{M}.

  11. end function

{λ1i,…,λMi}\{\lambda_{1i},\ldots,\lambda_{Mi}\} as λ(1)i≤λ(2)i≤⋯≤λ(M)i\lambda_{(1)i}\le\lambda_{(2)i}\le\cdots\le\lambda_{(M)i}, and then the upper bound of each interval is defined as

din:=λ(⌊nM/N⌋)i+λ(⌊nM/N⌋+1)i2,n=1,…,N−1.d_{in}:=\frac{\lambda_{(\lfloor nM/N\rfloor)i}+\lambda_{(\lfloor nM/N\rfloor+1)i}}{2},\qquad n=1,\ldots,N-1.

The first interval is Bi1:=(0,di1]\mathcal{B}_{i1}:=(0,d_{i1}], the last interval is BiN:=(di,N−1,∞)\mathcal{B}_{iN}:=(d_{i,N-1},\infty), and the intervals in between are Bin:=(di,n−1,din]\mathcal{B}_{in}:=(d_{i,n-1},d_{in}], n=2,…,N−1n=2,\ldots,N-1. For the observed λi\lambda_i, suppose it falls in Bin\mathcal{B}_{in}, where n∈{1,…,N}n\in\{1,\ldots,N\}, then the solved optimal surge staffing level is

s^in:=round⁡(NM∑m:λmi∈Binsmi(b^)).(13)\hat{s}_{in}:=\operatorname{round}\left(\frac{N}{M}\sum_{m:\lambda_{mi}\in\mathcal{B}_{in}}s_{mi}(\hat{b})\right). \tag*{(13)}

In the implementation, to balance the computation burden and accuracy, after obtaining b^\hat{b} and {sm(b^)}m=1M\{s_m(\hat{b})\}_{m=1}^{M} for some MM, we can generate additional samples of Λ\boldsymbol{\Lambda} for better arrival rate interval construction and surge staffing level aggregation. Say, for some M′>MM'>M, additionally generate λm\boldsymbol{\lambda}_m, m=M+1,…,M′m=M+1,\ldots,M' and compute the associated {sm(b^)}m=M+1M′\{s_m(\hat{b})\}_{m=M+1}^{M'} by calling Evaluate function in Algorithm 1. Then the above interval construction and solution aggregation are with respect to {λmi}m=1M′\{\lambda_{mi}\}_{m=1}^{M'} and {sm(b^)}m=1M′\{s_m(\hat{b})\}_{m=1}^{M'}. Note that the aforementioned method is only one possibility to realize the offline simulation online application idea. While more sophisticated methodologies such as surrogate model construction could certainly be employed, we adopt it due to its simplicity and satisfactory performance (which will be shown in numerical results).

Two-Stage Staffing Lookup Table

To be more conveniently used in practice, the estimated optimal surge staffing level s^in\hat{s}_{in} can be pre-computed for the possible values of λi\lambda_i before we actually observe it. Namely, for each shift ii, we only need to compute s^in\hat{s}_{in} for each interval Bin\mathcal{B}_{in}, n=1,…,Nn=1,\ldots,N, and record it. With this, we can make a two-stage staffing lookup table far before the successive shifts start. Table 1 shows a sketch of such a staffing lookup table. In practice, for adjacent intervals of λi\lambda_i, they can be merged together if their surge staffing levels are the same.

Shift ii12⋯\cdotsII
Solved base staffing b^i\hat{b}_ib^1\hat{b}_1b^2\hat{b}_2⋯\cdotsb^I\hat{b}_I
Solved surge staffing s^i\hat{s}_i for different intervals of λi\lambda_iB11↦s^11\mathcal{B}_{11} \mapsto\hat{s}_{11}B21↦s^21\mathcal{B}_{21} \mapsto\hat{s}_{21}⋯\cdotsBI1↦s^I1\mathcal{B}_{I1} \mapsto\hat{s}_{I1}
B12↦s^12\mathcal{B}_{12} \mapsto\hat{s}_{12}B22↦s^22\mathcal{B}_{22} \mapsto\hat{s}_{22}⋯\cdotsBI2↦s^I2\mathcal{B}_{I2} \mapsto\hat{s}_{I2}
⋮\vdots⋮\vdots⋱\ddots⋮\vdots
B1N↦s^1N\mathcal{B}_{1N} \mapsto\hat{s}_{1N}B2N↦s^2N\mathcal{B}_{2N} \mapsto\hat{s}_{2N}⋯\cdotsBIN↦s^IN\mathcal{B}_{IN} \mapsto\hat{s}_{IN}

Table 1. Sketch of a two-stage staffing lookup table based on SOA.

Since both the base staffing decision and surge staffing decision are computed well in advance, the substantial time required for simulation and optimization exerts negligible impact on real-time decision-making. Users only need to check the staffing lookup table to decide the base staffing levels for all shifts in the beginning and the surge staffing level of each shift ii upon observing λi\lambda_i (simply by finding the corresponding interval in the table that covers λi\lambda_i). The two layers of approximation on the model side (proposed in Section 3.2.1) and the compromise approach used in estimating s~i∗\tilde{s}_i^* (proposed in Section 3.2.2) of course make the final solutions (thus the two-stage staffing lookup table) sub-optimal. However, as will be demonstrated shortly in the numerical experiments, the quality of this SOA-based staffing policy closely approaches the true optimal solution and significantly outperforms existing approaches in the literature.

Comparison with Existing Policies

Note that Bassamboo et al. [3] and Hu et al. (2025) have proposed explicit staffing policies for more specific settings in EDs. We first summarize their setting and assumptions, then present their policies, along with necessary discussion on how our SOA is applied in their settings.

Their basic models share the same setting and assumptions. For shift ii, patient arrivals are assumed to follow a DSPP with random arrival rate Λi\Lambda_i, which has the following structure:

Λi:=λˉi+Xiλˉiα,(14)\Lambda_i := \bar{\lambda}_i + X_i\bar{\lambda}_i^\alpha, \tag*{(14)}

where Xi∼N(0,σ2)X_i \sim\mathcal{N}(0, \sigma^2), α∈(0,1)\alpha\in(0, 1) is a constant, and λˉi>0\bar{\lambda}_i > 0 is a constant for each ii. Let FiF_i denote the distribution of Λi\Lambda_i, and it is easy to see that E[Λi]=λˉi\mathbb{E}[\Lambda_i] = \bar{\lambda}_i. Note that the support of Λi\Lambda_i in (14) contains negative values. In practice, only the positive samples of Λi\Lambda_i are retained (i.e., the truncated version of Λi\Lambda_i is actually used). Conditional on Λi=λi\Lambda_i = \lambda_i, the arrival process is a homogeneous Poisson process, and the ED is assumed to operate as an M/M/k+MM/M/k+M queue. The service and patience times are independent exponential random variables with means 1/μ1/\mu and 1/θ1/\theta, respectively, and are independent of the arrival process. For each individual in the queue, when the waiting time reaches the patience time, s/he will leave the system immediately. The initial condition is left unspecified, as they both treat the service system as if it operates in steady state, whereas the underlying problem inherently spans a finite horizon with multiple shift transitions.

The considered QoS includes the queue length and abandonment rate. To be precise, the cost function over II consecutive shifts is defined as follows. For notational convenience, normalize the shift length to one (by adjusting all the rates), thus the total length for II consecutive shifts is II. Let Q(t;k,λ,π0)Q(t;\mathbf{k},\boldsymbol{\lambda},\pi_0) denote the queue length at time tt, and B(t;k,λ,π0)B(t;\mathbf{k},\boldsymbol{\lambda},\pi_0) the cumulative number of abandonments up to time tt, where t∈[0,I]t \in[0,I]. Then, the expected penalty related to the QoS of the service system is

g(k,λ,π0)=hE[∫0IQ(t;k,λ,π0) dt]+aE[B(I;k,λ,π0)].g(\mathbf{k},\boldsymbol{\lambda},\pi_0)=h\mathbb{E}\left[\int_0^I Q(t;\mathbf{k},\boldsymbol{\lambda},\pi_0)\,\mathrm{d}t\right]+a\mathbb{E}[B(I;\mathbf{k},\boldsymbol{\lambda},\pi_0)].

where expectations are taken with respect to the queueing randomness, hh is the waiting cost per customer per unit time and aa is the cost per abandonment. However, applying their formulations separately to each shift gives

g(k,λ)=(h+aθ)∑i=1Iq(ki,λi).g(\mathbf{k},\boldsymbol{\lambda})=(h+a\theta)\sum_{i=1}^I q(k_i,\lambda_i).

where q(ki,λi)q(k_i,\lambda_i) denotes the expectation of the steady-state queue length, as they treat the system as if it operates in steady state for each shift. This is certainly a simplifying approximation, which constitutes one of the reasons for the sub-optimal performance of their policies in practice.

Based on such assumptions of the stylized models of arrival and service system, and the steady-state approximation of the QoS, Bassamboo et al. [3] propose a single-stage newsvendor policy in light of a stochastic-fluid approximation, which is denoted as P1,NV\mathcal{P}_{1,\mathrm{NV}} in this paper. Formally, P1,NV\mathcal{P}_{1,\mathrm{NV}} specifies that

bi1,NV=1μFi−1(1−ci(b)hμ/θ+aμ).b_i^{1,\mathrm{NV}}=\frac{1}{\mu}F_i^{-1}\left(1-\frac{c_i^{(b)}}{h\mu/\theta+a\mu}\right).

and si1,NV=0s_i^{1,\mathrm{NV}}=0. Under the arrival-rate scaling and regularity assumptions, P1,NV\mathcal{P}_{1,\mathrm{NV}} is proved to have an optimality gap of O(λˉi1−α)O(\bar{\lambda}_i^{1-\alpha}) relative to the optimal single-stage staffing cost for α∈(1/2,1)\alpha\in(1/2,1), as λˉi→∞\bar{\lambda}_i\to\infty ([3]). Hu et al. (2025) propose a two-stage QED policy by combining the fluid-based staffing rule with a diffusion approximation to account for queueing variability, which is denoted as P2,QED\mathcal{P}_{2,\mathrm{QED}} in this paper. Formally, for the regime hμ/θ+aμ>ci(s)>ci(b)h\mu/\theta+a\mu>c_i^{(s)}>c_i^{(b)}, P2,QED\mathcal{P}_{2,\mathrm{QED}} specifies that

bi2,QED=1μFi−1(1−ci(b)ci(s))+ηi∗λˉiμ,b_i^{2,\mathrm{QED}}=\frac{1}{\mu}F_i^{-1}\left(1-\frac{c_i^{(b)}}{c_i^{(s)}}\right)+\eta_i^*\sqrt{\frac{\bar{\lambda}_i}{\mu}},

and, after observing λi\lambda_i,

si2,QED(λi)=(λiμ+ηi∗λiμ−bi2,QED)+,s_i^{2,\mathrm{QED}}(\lambda_i)=\left(\frac{\lambda_i}{\mu}+\eta_i^*\sqrt{\frac{\lambda_i}{\mu}}-b_i^{2,\mathrm{QED}}\right)^+,

where (x)+=max⁡{x,0}(x)^+=\max\{x,0\}, and the parameter ηi∗\eta_i^* is obtained by solving

ηi∗=arg⁡min⁡η∈R{ci(s)η+(hμθ+aμ)θ/μ[H(ημ/θ)−ημ/θ]1+θ/μH(ημ/θ)/H(−η)},\eta_i^*=\arg\min_{\eta\in\mathbb{R}}\left\{c_i^{(s)}\eta+\left(\frac{h\mu}{\theta}+a\mu\right)\frac{\sqrt{\theta/\mu}\left[H\left(\eta\sqrt{\mu/\theta}\right)-\eta\sqrt{\mu/\theta}\right]}{1+\sqrt{\theta/\mu}H\left(\eta\sqrt{\mu/\theta}\right)/H(-\eta)}\right\},

where H(x)=ϕ(x)/Φ(−x)H(x)=\phi(x)/\Phi(-x) is the standard normal hazard function, and ϕ\phi and Φ\Phi denote the standard normal density and distribution functions, respectively. Under certain assumptions, P2,QED\mathcal{P}_{2,\mathrm{QED}} is proved to have an optimality gap of o(λˉi)o(\sqrt{\bar{\lambda}_i}) relative to the optimal two-stage staffing cost for α∈(1/2,1)\alpha\in(1/2,1), as λˉi→∞\bar{\lambda}_i\to\infty (Hu et al., 2025). Both policies need to round their staffing decisions to integers when deployed in practice. For the multi-shift case, Hu et al. (2025) suggest some heuristic adjustments for compensating the transient-shift effect:

biadj=bi2,QED+ξ1Δi,siadj=si2,QED(λi)+ξ2Di,b_i^{\mathrm{adj}}=b_i^{2,\mathrm{QED}}+\xi_1\Delta_i,\qquad s_i^{\mathrm{adj}}=s_i^{2,\mathrm{QED}}(\lambda_i)+\xi_2D_i,

where Δi\Delta_i is the difference in expected steady-state queue lengths between the preceding and current shifts, and DiD_i measures the deviation of the initial queue length from its steady-state expectation under the current staffing. The parameters ξ1\xi_1 and ξ2\xi_2 control the base and surge adjustments, respectively.

Numerical Experiments

Before presenting the experimental setup, we must provide a supplementary description of the operational rules of the service system under study. When the staffing level kik_i varies among shifts, the service system must have a clear rule to handle the “customers” in the “servers” if the number of servers is reduced in a subsequent shift. Throughout this paper, we consider the same simple rule as in Hu et al. (2025). When the number of servers is about to be reduced in the beginning of a shift, all “customers” in the “servers” have to suspend their service and rejoin the front of the waiting line, where an earlier arrival time corresponds to a more advanced position in the queue. When they re-enter the server, their service can resume from where it was interrupted.

It is worth emphasizing that how the system handles entities currently in service when the number of servers decreases is merely an operational policy. While this rule dictates the low-level execution of the simulation for performance estimation, it remains completely independent of our proposed simulation-based optimization framework. This issue has rarely been mentioned in existing studies, because they neither conduct a rigorous analytical derivation nor a strict simulation for the multi-shift case.

We use CRNs to compare candidate staffing decisions. For each arrival rate realization, all candidates share the same arrival, service-time, and patience-time samples within each replication to reduce simulation noise in cost comparisons. To avoid repeated simulation, we cache the estimated waiting and abandonment costs for each arrival rate and total staffing configuration. With the initial state and random samples unchanged, different base-surge combinations with the same total staffing reuse these estimates, while their staffing costs are calculated separately. All numerical experiments are implemented in Python on a personal computer equipped with an Intel Core i5-11400 processor (2.60 GHz) and 32 GB of RAM. Source code is available at https://github.com/shenhaihui/soa.

Single-Shift Case

We consider two versions for our proposed SOA. The first version is to directly solve bˉi∗\bar{b}_i^* in (7) and si∗s_i^* in (8), which is denoted as P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}. Note that for the single-shift case, P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} closely approaches the true optimum (up to the errors in simulation and optimization), because no approximations are introduced into the original formulation. The second version is to construct the two-stage staffing lookup table by solving bˉi∗\bar{b}_i^* in (7) and s^in\hat{s}_{in} in (13), which is denoted as P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. For comparison, we also implement P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}}, which are discussed in § 3.3. Moreover, to better investigate the benefit of SOA, we also consider two variants. The first one is called P1,SOA\mathcal{P}_{1,\mathrm{SOA}}, which solves bˉi∗\bar{b}_i^* in (7) by restricting si=0s_i = 0. It means that we only consider the base staffing in SOA.

The second one is called PQED+SOA\mathcal{P}_{\mathrm{QED+SOA}}, where the base staffing decision is made via P2,QED\mathcal{P}_{2,\mathrm{QED}} and the surge staffing decision is made via P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} by directly solving sˉi∗\bar{s}_i^* in (8). It is used to individually investigate the benefit of SOA in the first-stage decision (when compared to P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} or P2,SOA\mathcal{P}_{2,\mathrm{SOA}}) and the benefit of SOA in the second-stage decision (when compared to P2,QED\mathcal{P}_{2,\mathrm{QED}}).

Special Problem Setting

We first consider the special problem setting of nurse staffing in EDs in Hu et al. (2025) (as presented in § 3.3). Since Hu et al. (2025) treat the system as if it operates in steady state, the initial condition is not clearly specified there. We will try different initial conditions.

Experiment 1. To begin with, we let the system start empty and idle. Other problem parameters are exactly the same as in Hu et al. (2025). That is, the parameters in the random arrival rate (14) are λˉi=100\bar{\lambda}_i = 100, α=0.75\alpha= 0.75 and σ=1\sigma= 1. The parameters for the queue are μ=1\mu= 1 and θ=0.1\theta= 0.1, and the cost parameters are h=1.5h = 1.5, a=3a = 3, ci(b)=1c_i^{(b)} = 1, and ci(s)=1.5c_i^{(s)} = 1.5. In the implementation of SOA, M=M′=1,000M = M' = 1{,}000, R=100R = 100, and N=20N = 20. In P1,SOA\mathcal{P}_{1,\mathrm{SOA}}, the search of base staffing solution starts from the solution of P1,NV\mathcal{P}_{1,\mathrm{NV}}; in P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, the search of base staffing solution and surge staffing solution starts from the solutions of P2,QED\mathcal{P}_{2,\mathrm{QED}} (hereinafter the same). For evaluation, we re-generate 1,000 arrival rates and run 100 simulation replications of the system under each scenario, independently of the offline search. For P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}, the surge staffing is solved online with another 100 simulation replications of the system.

Table 2 shows the performance comparison among six policies, and Table 3 shows the constructed staffing lookup table of P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. Several key findings are as follows. (1) The average total costs under P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} are almost the same, indicating that constructing the lookup table offline as a substitute of real-time simulation optimization for surge staffing decision incurs negligible optimality loss. (2) The average total cost under P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is significantly lower than that under P1,NV\mathcal{P}_{1,\mathrm{NV}} (by 61.8%) and that under P2,QED\mathcal{P}_{2,\mathrm{QED}} (by 54.2%). (3) Comparing P1,SOA\mathcal{P}_{1,\mathrm{SOA}} to P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} or P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, for this specific problem instance, the second-stage decision plays a marginal role because the corresponding surge staffing level is negligible compared to the base staffing level. The underlying reason is that the variance of the random arrival rate here is relatively small, which will be further substantiated in subsequent experiments. (4) Nevertheless, the single-stage policy

MetricP2,SOA−DP_{2,\mathrm{SOA-D}}P2,SOAP_{2,\mathrm{SOA}}P1,NVP_{1,\mathrm{NV}}P2,QEDP_{2,\mathrm{QED}}P1,SOAP_{1,\mathrm{SOA}}PQED+SOAP_{\mathrm{QED}+\mathrm{SOA}}
Average total cost57.29057.288150.000125.21857.34094.192
Base staffing bib_i3838150943994
Average surge staffing sis_i1.1481.2490.00020.8120.0000.000
Offline search time (s)2,1152,115––1,569–
Online search time (s)1,079––––391

Table 2. Performance comparison of six policies for Experiment 1.

Shift ii1
Solved base staffing b^i\hat{b}_i38
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(0,122.295]↦0(0, 122.295] \mapsto0
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(122.295,127.224]↦1(122.295, 127.224] \mapsto1
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(127.224,133.160]↦2(127.224, 133.160] \mapsto2
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(133.160,143.488]↦4(133.160, 143.488] \mapsto4
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(143.488,155.819]↦8(143.488, 155.819] \mapsto8
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(155.819,∞)↦14(155.819,\infty) \mapsto14
Note: Adjacent intervals with the same surge staffing are merged.

Table 3. Two-stage staffing lookup table of P2,SOAP_{2,\mathrm{SOA}} for the single-shift case in Experiment 1.

P1,NVP_{1,\mathrm{NV}} performs the worst because it yields an excessive base staffing level. (5) P2,QEDP_{2,\mathrm{QED}} also yields an inflated base staffing level—albeit slightly lower than that of P1,NVP_{1,\mathrm{NV}}—but further introduces a substantial surge staffing level, pushing its total staffing level close to that of P1,NVP_{1,\mathrm{NV}}. This is because they are derived by treating the system as operating in a steady state, whereas in reality the system starts empty. Consequently, the actual resource requirement during the transient phase is substantially lower than expected. Subsequent experiments will further substantiate this point.

(6) Comparing PQED+SOAP_{\mathrm{QED}+\mathrm{SOA}} to P2,SOA−DP_{2,\mathrm{SOA-D}} (or P2,SOAP_{2,\mathrm{SOA}}) and P2,QEDP_{2,\mathrm{QED}}, it is clear that both the first- and second-stage decisions in the SOA framework have unique and irreplaceable advantages.

Experiment 2. We examine the number of patients remaining in the system at the end of the shift. Across all policies and simulation replications, the average is approximately 67, with the 5th percentile at around 29 and the 95th percentile at around 110. In this regard, we repeat Experiment 1 by initializing the system with 30, 60, 90, and 120 unfinished patients, respectively. It is worth noting that the decisions under P1,NVP_{1,\mathrm{NV}} and P2,QEDP_{2,\mathrm{QED}} are unaffected by these variations, as they do not depend on the initial state. Hence, the base staffing decision in PQED+SOAP_{\mathrm{QED}+\mathrm{SOA}} also persists.

Table 4 shows the comparisons of costs and staffing decisions among six policies over different initial states. Several key findings are as follows. (1) The performances of P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} are still close and dominantly better than other policies. As the initial number of patients increases, their associated base staffing levels increase, and the total costs also increase. (2) For P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}}, the average total costs are relatively insensitive to initial conditions, as their massive staffing levels buffer the QoS against moderate changes in the initial number of patients. (3) Comparing P2,QED\mathcal{P}_{2,\mathrm{QED}} with P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, the cost savings delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} gradually shrink from 54.2% to 10.5% as the initial number of patients increases from 0 to 120, demonstrating increasing performance of P2,QED\mathcal{P}_{2,\mathrm{QED}}. This aligns with the fact that P2,QED\mathcal{P}_{2,\mathrm{QED}} is derived under the assumption of steady-state operation. A relatively heavier initial workload accelerates the system’s convergence toward steady state within a finite horizon, which enhances the validity of the approximation in P2,QED\mathcal{P}_{2,\mathrm{QED}}, resulting in better performance. A similar trend can be observed for P1,NV\mathcal{P}_{1,\mathrm{NV}}. (4) Other observations are the same as in Experiment 1.

InitializationP2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}P2,SOA\mathcal{P}_{2,\mathrm{SOA}}P1,NV\mathcal{P}_{1,\mathrm{NV}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P1,SOA\mathcal{P}_{1,\mathrm{SOA}}PQED+SOA\mathcal{P}_{\mathrm{QED}+\mathrm{SOA}}
3072.78872.785150.000125.21872.96894.799
6088.77788.770150.002125.22289.22997.094
90105.797105.797150.022125.608106.613105.875
120125.163125.166150.160139.902126.063131.723
InitializationP2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}P2,SOA\mathcal{P}_{2,\mathrm{SOA}}P1,NV\mathcal{P}_{1,\mathrm{NV}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P1,SOA\mathcal{P}_{1,\mathrm{SOA}}PQED+SOA\mathcal{P}_{\mathrm{QED}+\mathrm{SOA}}
30(55, 2.030)(55, 2.130)(150, 0.000)(94, 20.812)(56, 0.000)(94, 0.000)
60(73, 2.849)(73, 2.884)(150, 0.000)(94, 20.812)(75, 0.000)(94, 0.034)
90(92, 3.384)(92, 3.488)(150, 0.000)(94, 20.812)(95, 0.000)(94, 2.541)
120(115, 2.541)(115, 2.546)(150, 0.000)(94, 20.812)(118, 0.000)(94, 17.498)

Table 4. Cost and staffing comparisons of six policies for Experiment 2.

Experiment 3. We also try some different values for the parameters in the arrival process and service system. For two different initial states (60 and 120 unfinished patients, respectively), we vary the holding cost h∈{1.5,20}h \in\{1.5, 20\} with a=2ha = 2h, and the perturbation variance σ2∈{1,3.5}\sigma^2 \in\{1, 3.5\}. For the new parameter combinations that were not examined in Experiment 2, we rerun the experiments.

Initializationhhσ2\sigma^2P2,SOA\mathcal{P}_{2,\mathrm{SOA}}P1,NV\mathcal{P}_{1,\mathrm{NV}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P1,SOA\mathcal{P}_{1,\mathrm{SOA}}
601.51.088.770150.002125.22289.229
60201.0107.215183.000136.746118.426
601.53.596.774194.004143.73798.557
60203.5118.640256.000155.985145.775
1201.51.0125.166150.160139.902126.063
120201.0137.720183.006197.404147.794
1201.53.5132.539194.076162.362135.146
120203.5148.401256.000262.536174.845

Table 5. Cost comparison of four policies for Experiment 3.

Table 5 shows the numerical results. All the previous observations remain. Moreover, we can see that for a given hh, an increase in σ2\sigma^2 further amplifies the advantage of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}}. For example, for 120 initial unfinished patients, h=1.5h = 1.5 and σ2=1\sigma^2 = 1, the cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P2,QED\mathcal{P}_{2,\mathrm{QED}} is 10.5%, and it increases to 18.4% when σ2=3.5\sigma^2 = 3.5. It reveals that P2,SOA\mathcal{P}_{2,\mathrm{SOA}} can better cover the larger variance of arrival rates. When h=20h = 20 and σ2=3.5\sigma^2 = 3.5, the cost saving is even higher, which is 43.5%. The entire results demonstrate the high robustness of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} among different situations. Moreover, comparing P2,SOA\mathcal{P}_{2,\mathrm{SOA}} with P1,SOA\mathcal{P}_{1,\mathrm{SOA}} also reveals a similar fact. Note that in Experiment 1 we find the second-stage decision of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} plays a marginal role, where σ2=1\sigma^2 = 1. In this experiment, when σ2=1\sigma^2 = 1, the costs of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} and P1,SOA\mathcal{P}_{1,\mathrm{SOA}} are still very close. When σ2=3.5\sigma^2 = 3.5, the advantage of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,SOA\mathcal{P}_{1,\mathrm{SOA}} may become clear. For example, for 60 initial unfinished patients, h=20h = 20 and σ2=3.5\sigma^2 = 3.5, the cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is 18.6%.

General Problem Setting

Experiment 4. We now consider a situation where the arrival process and service system depart from the stylized models assumed in Bassamboo et al. [3] and Hu et al. (2025). Note that substantial structural changes—such as drastically modifying the QoS metrics or the associated cost functions—would cause analytical policies like P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}} to conceptually break down and lose their mathematical validity. So here we only consider localized variations. Suppose the interarrival times follow a Gamma distribution with mean 1/λi1/\lambda_i and squared coefficient of variation (SCV) 2, given that the rate is λi\lambda_i. The service times follow a lognormal distribution with mean 1/μ1/\mu and SCV 2. Moreover, there is balking behavior in the queue, i.e., an arrival leaves immediately when the number of waiting patients exceeds a threshold BB, and no cost is counted. Let B=20B = 20, and the values of other parameters remain the same as in Experiment 1. We call this setting a non-stylized model and the setting in Experiment 1 a stylized model.

P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}} will make the same staffing decisions as in Experiment 1, because they are forced to treat the system strictly as a stylized model, and the equivalent rates remain identical. For SOA, the difference in the setting can be well captured in the simulation implementation. Table 6 shows the numerical results for two different initial states (60 and 120 unfinished patients, respectively). It can be seen that when the actual problem deviates from the stylized model assumed in Bassamboo et al. [3] and Hu et al. (2025), the advantage of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}} may further amplify. For example, for 60 initial unfinished patients, the cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P2,QED\mathcal{P}_{2,\mathrm{QED}} increases from 29.1% to 35.7% when the actual problem changes from stylized model to non-stylized model. The reason is that analytical policies derived under stylized models, such as P2,QED\mathcal{P}_{2,\mathrm{QED}}, can not fully capture the key characteristics of the problem. It can be inferred that if we consider more substantial structural changes in the settings, the performance gap could be widened even further.

InitializationModelAverage total cost\mathop{\mathrm{Average\ total\ cost}}
P2,SOA\mathcal{P}_{2,\mathrm{SOA}}P1,NV\mathcal{P}_{1,\mathrm{NV}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P1,SOA\mathcal{P}_{1,\mathrm{SOA}}
60Stylized88.770150.002125.22289.229
Non-stylized80.570150.002125.22380.544
120Stylized125.166150.160139.902126.063
Non-stylized109.361150.066133.186109.361
InitializationModelStaffing(bi,Avg. si)\mathop{\mathrm{Staffing}}(b_i,\mathrm{Avg.}\ s_i)
P2,SOA\mathcal{P}_{2,\mathrm{SOA}}P1,NV\mathcal{P}_{1,\mathrm{NV}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P1,SOA\mathcal{P}_{1,\mathrm{SOA}}
60Stylized(73, 2.884)(150, 0.000)(94, 20.812)(75, 0.000)
Non-stylized(62, 0.268)(150, 0.000)(94, 20.812)(62, 0.000)
120Stylized(115, 2.546)(150, 0.000)(94, 20.812)(118, 0.000)
Non-stylized(69, 0.000)(150, 0.000)(94, 20.812)(69, 0.000)

Table 6. Cost and staffing comparisons of four policies for Experiment 4.

Multi-Shift Case

As in Hu et al. (2025), for the multi-shift case, correlation between arrival rates in consecutive shifts is considered, where the XiX_i’s in (14) satisfy

X1=σZ1,Xi=ρXi−1+σ1−ρ2Zi,i=2,…,I,(15)X_1 = \sigma Z_1,\qquad X_i = \rho X_{i-1} + \sigma\sqrt{1-\rho^2}Z_i,\qquad i=2,\ldots,I, \tag*{(15)}

where Z1,…,ZIZ_1,\ldots,Z_I are independent standard normal random variables, and ρ\rho controls the correlation between consecutive shifts. In contrast to the single-shift case, the SOA framework for the general multi-shift case relies additionally on two layers of approximation on the model side. We first try to investigate the “optimality gap” of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} in a very small-scale problem. Then, on the case study of nurse staffing with real data investigated in Hu et al. (2025), P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is compared with other policies.

Small Scale

For very small problem size, it is possible to directly solve the original optimization problem, i.e., b∗b^* in (4) and si∗s_i^* in (5), using simulation optimization. Such a solution is treated as the “true optimal” solution (of course subject to simulation optimization error), and we use P2,OPT\mathcal{P}_{2,\mathrm{OPT}} to denote the corresponding policy. We use P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} to denote the situation that b~ˉ∗\bar{\tilde{b}}^* in (12) is solved for base staffing, and s~i∗\tilde{s}_i^* in (11) is solved for surge staffing, which means we re-generate samples of Λ\Lambda conditional on Λi=λi\Lambda_i=\lambda_i and do not use the offline-generated sample paths via the lookup table. It is, of course, impossible for (near-)real-time decision in the general multi-shift case. We use P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} to see the optimality gap due to the approximation on the model side. The practical P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is also applied to investigate the overall optimality gap.

Experiment 5. Consider the smallest multi-shift problem, i.e., I=2I=2. It is the simple extension of the problem in Experiment 1, with ρ=0.3\rho=0.3 in (15) representing a positive correlation, and all other parameters the same as in Experiment 1. The system starts empty, and patients at the end of shift 1 carry over to shift 2. In the implementation of P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, M=100M=100, M′=1000M'=1000, R=20R=20, N=20N=20, Kout=1K_{\mathrm{out}}=1 and Kin=2K_{\mathrm{in}}=2. For evaluation, we re-generate 100 arrival rate vectors and run 100 simulation replications of the system under each scenario, independently of the offline search.

P2,OPT\mathcal{P}_{2,\mathrm{OPT}} is implemented as a standard dynamic programming procedure, where the objective function evaluation is simulation-based, and the solution search is powered by the same coordinate-search algorithm as in P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. For P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}, to solve the surge staffing online, for each re-generated arrival rate vector used in evaluation, additional conditional arrival rate vectors are generated for the conditional expectation estimation in (11). More details on the implementation can be found in the open-source code.

Table 7 shows the optimality gap and computation time and Table 8 shows the constructed staffing lookup table of P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. It can be seen that, at least for this two-shift problem, the two

PolicyTotal cost (95% CI)(b1,b2)(b_1,b_2)(s‾1,s‾2)(\overline{s}_1,\overline{s}_2)Offline time (s)Online time (s)
P2,OPT\mathcal{P}_{2,\mathrm{OPT}}159.233±8.073159.233 \pm8.073(42,78)(42,78)(4.470,9.191)(4.470,9.191)17,19611,063
P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}}164.643±8.563164.643 \pm8.563(42,79)(42,79)(3.500,9.070)(3.500,9.070)457724
P2,SOA\mathcal{P}_{2,\mathrm{SOA}}164.135±8.607164.135 \pm8.607(42,79)(42,79)(3.510,8.140)(3.510,8.140)848–

Table 7. Optimality gap and computation time for Experiment 5. Note: Offline times include base staffing search and, for P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, lookup table construction.

Shift ii12
Solved base staffing b^i\hat{b}_i4279
Solved surge staffing s^i\hat{s}_i, for different intervals of λi\lambda_i(0,112.370]↦0(0,112.370] \mapsto0(0,74.362]↦0(0,74.362] \mapsto0
(112.370,117.478]↦1(112.370,117.478] \mapsto1(74.362,79.221]↦1(74.362,79.221] \mapsto1
(117.478,121.664]↦3(117.478,121.664] \mapsto3(79.221,83.183]↦2(79.221,83.183] \mapsto2
(121.664,126.947]↦4(121.664,126.947] \mapsto4(83.183,87.802]↦3(83.183,87.802] \mapsto3
(126.947,132.199]↦6(126.947,132.199] \mapsto6(87.802,95.596]↦4(87.802,95.596] \mapsto4
(132.199,138.871]↦9(132.199,138.871] \mapsto9(95.596,103.520]↦5(95.596,103.520] \mapsto5
(138.871,153.504]↦13(138.871,153.504] \mapsto13(103.520,106.868]↦6(103.520,106.868] \mapsto6
(153.504,∞)↦24(153.504,\infty) \mapsto24(106.868,110.478]↦7(106.868,110.478] \mapsto7
(110.478,115.249]↦8(110.478,115.249] \mapsto8
(115.249,120.393]↦6(115.249,120.393] \mapsto6
(120.393,125.745]↦14(120.393,125.745] \mapsto14
(125.745,131.740]↦16(125.745,131.740] \mapsto16
(131.740,138.190]↦17(131.740,138.190] \mapsto17
(138.190,149.118]↦19(138.190,149.118] \mapsto19
(149.118,∞)↦28(149.118,\infty) \mapsto28

Table 8. Two-stage staffing lookup table of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} for the two-shift case in Experiment 5. Note: Adjacent intervals with the same surge staffing are merged.

layers of model-side approximations are tight, as the performance gap between P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} and the “true optimal” policy P2,OPT\mathcal{P}_{2,\mathrm{OPT}} is only around 3% in mean. Furthermore, consistent with the findings in the single-shift case, constructing the lookup table offline as a substitute for real-time simulation-optimization for surge staffing decisions incurs a negligible loss of optimality, as the performances of P2,SOA−D\mathcal{P}_{2,\mathrm{SOA-D}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} are indistinguishable. Note that the computation time for P2,OPT\mathcal{P}_{2,\mathrm{OPT}} is substantially longer than that for P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} effectively shifts the online computational burden to offline. Also note that in the constructed staffing lookup table, due to the existence of errors, the surge staffing does not strictly increase with realized arrival rates, which is not reasonable. Two refinement directions exist: either appending additional offline simulations around anomalous results to improve local precision, or applying a rule-based post-processing step to enforce monotonicity. We do not elaborate further on these methods in the present paper.

Practical Case in Hu et al. (2025)

Now we consider the practical one-week ED staffing problem with real data, investigated in Hu et al. (2025). It consists of I=14I = 14 consecutive 12-hour shifts, and for convenience the shift length is normalized to one. The expected arrival rate for each shift, λˉi\bar{\lambda}_i, is as shown in Table 9. The arrival process is assumed to satisfy the models (14) and (15) with parameters α=0.769\alpha= 0.769, σ2=0.118\sigma^2 = 0.118 and ρ=0.3\rho= 0.3. The ED is assumed to operate as an M/M/k+MM/M/k + M queue given the realized arrival rate, with μ=1.4\mu= 1.4 and θ=12/27.5\theta= 12/27.5. The staffing cost parameters are ci(b)=540c_i^{(b)} = 540 and ci(s)=810c_i^{(s)} = 810 (corresponding to hourly wages of $45 and $67.5, respectively), and the waiting and abandonment cost parameters are h=810h = 810 and a=1620a = 1620. The system starts empty, and unfinished patients carry over across shifts.

SunMonTueWedThuFriSat
Day λˉi\bar{\lambda}_i141.019207.385188.769186.942185.208175.173147.058
Night λˉi\bar{\lambda}_i89.46297.05897.76993.71195.18996.69294.115

Table 9. Shift-level mean arrival rates based on the ED data (Hu et al., 2025).

We compare P1,NV\mathcal{P}_{1,\mathrm{NV}}, P2,QED\mathcal{P}_{2,\mathrm{QED}}, P1,SOA\mathcal{P}_{1,\mathrm{SOA}}, and P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. For P2,QED\mathcal{P}_{2,\mathrm{QED}}, the suggested heuristic adjustments in Hu et al. (2025) as described in § 3.3 are applied, where ξ1=5\xi_1 = 5 and ξ2=1\xi_2 = 1. In the implementation of P2,SOA\mathcal{P}_{2,\mathrm{SOA}}, M=100M = 100, M′=4000M' = 4000, R=10R = 10, N=20N = 20, Kout=1K_{\mathrm{out}} = 1 and Kin=2K_{\mathrm{in}} = 2. We have

tried larger KoutK_{\mathrm{out}} and found that the performance improvement is marginal (see Appendix § 6.1 for details). For evaluation, we re-generate 100 arrival rate vectors and run 100 simulation replications of the system under each scenario, independently of the offline search.

Table 10 shows the performance of four policies, and the full operational two-stage staffing lookup table constructed and used in P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is collected in Appendix § 6.2. We can see that, in terms of the total cost, P2,SOA\mathcal{P}_{2,\mathrm{SOA}} performs the best. The cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is 19.96% over P1,NV\mathcal{P}_{1,\mathrm{NV}} in mean, and 7.87% over P2,QED\mathcal{P}_{2,\mathrm{QED}} in mean. We can see that P2,QED\mathcal{P}_{2,\mathrm{QED}} yields an excessive surge staffing level, which makes its total staffing level the highest. Although the resulting QoS is the best for all metrics, it does not trade off well with the staffing cost, which can be more clearly seen in Figure 1.

Comparison of average total cost and cost components under four policies

Figure 1. Comparison of average total cost and cost components under four policies.

Another interesting thing is, the one-stage policy P1,SOA\mathcal{P}_{1,\mathrm{SOA}} has very close cost performance compared with the two-stage policy P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. However, upon closer inspection, they differ in several fine details. First, P2,SOA\mathcal{P}_{2,\mathrm{SOA}} does decide some surge staffing levels, unlike P1,SOA\mathcal{P}_{1,\mathrm{SOA}}. Second, although

P1,SOA\mathcal{P}_{1,\mathrm{SOA}} yields a higher total staffing level, the resulting QoS is relatively close to that in P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. Of course, one should note that QoS is not only determined by the total staffing level, but also by the detailed configuration among different shifts. As a side note, P1,NV\mathcal{P}_{1,\mathrm{NV}} yields an even slightly higher total staffing level compared to P1,SOA\mathcal{P}_{1,\mathrm{SOA}}, but its resulting QoS is much worse. Third, as one surge staff costs more than one base staff, in terms of the total cost, P1,SOA\mathcal{P}_{1,\mathrm{SOA}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} exhibit no substantial difference.

We further examine the policy performance under different levels of demand uncertainty and correlation between consecutive shifts. Specifically, we consider σ2∈{0.118,0.236,0.354}\sigma^{2} \in\{0.118, 0.236, 0.354\} and ρ∈{0,0.3}\rho\in\{0, 0.3\} and repeat the experiments, while keeping all other parameters unchanged. Table 11 reports the mean total costs of three policies and the cost savings achieved by P2,SOA\mathcal{P}_{2,\mathrm{SOA}}. It can be seen that an increase in demand uncertainty (i.e., in σ2\sigma^{2}) will further amplify the advantage of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P2,QED\mathcal{P}_{2,\mathrm{QED}}. Specifically, the cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P2,QED\mathcal{P}_{2,\mathrm{QED}} increases from about 8% to about 13% when σ2\sigma^{2} increases from 0.118 to 0.354. This effect and its underlying reason are aligned with those in the single-shift case considered in Experiment 3. However, surprisingly, comparing P2,SOA\mathcal{P}_{2,\mathrm{SOA}} with P1,SOA\mathcal{P}_{1,\mathrm{SOA}}, the effect is not that strong, and the cost saving delivered by P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,SOA\mathcal{P}_{1,\mathrm{SOA}} increases slowly with σ2\sigma^{2}. Since this case study is based on real data, this observation reveals that in real applications the two-stage staffing framework may not always exhibit a pronounced advantage relative to the one-stage alternative, especially considering its added complexity in computation and execution. Recall that in the single-shift case considered in Experiment 3, we do see situations where P2,SOA\mathcal{P}_{2,\mathrm{SOA}} substantially outperforms P1,SOA\mathcal{P}_{1,\mathrm{SOA}}. Thus, we are not concluding that the two-stage staffing framework is universally redundant; rather, we offer a caveat that this more sophisticated

σ2\sigma^{2}ρ\rhoAverage total costSaving of P2,SOA\mathcal{P}_{2,\mathrm{SOA}}
P1,SOA\mathcal{P}_{1,\mathrm{SOA}}P2,QED\mathcal{P}_{2,\mathrm{QED}}P2,SOA\mathcal{P}_{2,\mathrm{SOA}}vs. P1,SOA\mathcal{P}_{1,\mathrm{SOA}}vs. P2,QED\mathcal{P}_{2,\mathrm{QED}}
0.1180.0870,197944,951868,0400.25%8.14%
0.1180.3875,056944,474870,1780.56%7.87%
0.2360.0889,599960,977884,2130.61%7.99%
0.2360.3898,385962,254888,1491.14%7.70%
0.3540.0907,1791,032,099898,6800.94%12.93%
0.3540.3919,1381,033,761903,5031.70%12.60%

Table 11. Comparison of average total cost with different demand variances and correlations.

staffing approach may not always be necessary in practice.

Finally, we consider the most realistic scenario, where an ED actually operates continuously week after week without resetting, although staffing schedules are planned weekly. Even over a time horizon free of significant seasonality, meaning that the mean arrival rate profile remains identical across weeks, unfinished patients at the end of one week carry over to the next week and the demand correlations persist across consecutive weeks. Admittedly, P2,SOA\mathcal{P}_{2,\mathrm{SOA}} is pre-computed offline under the one-week setting, and the two-stage staffing lookup table depends solely on the distributions of Λi\Lambda_i and the specified initial state at the beginning of the week, whereas real-time states are not directly accounted for in surge staffing decisions. However, the initial state at the beginning of each week is a consequence of previous operations and cannot be known a priori. By assuming the same initial states and ignoring the demand correlations across consecutive weeks, the two-stage staffing lookup table of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} remains identical across weeks with the same mean arrival rate profiles. A natural question is whether applying such a one-week policy cyclically in a long horizon remains effective.

To answer this question, we consider a continuous 52-week (728-shift) horizon under the baseline parameter setting (σ2=0.118\sigma^2 = 0.118, ρ=0.3\rho= 0.3), assuming that the arrival-rate distribution remains identical across weeks (whose means are as shown in Table 9) and the demand correlation (as modeled in (15)) evolves continuously across weeks. As previously noted, because unfinished patients carry over, each week is unlikely to start empty. To accommodate this issue, we use the average number of unfinished patients remaining at the end of shift 14 (which is 75), obtained from the previous baseline one-week experiment under the benchmark policy P2,QED\mathcal{P}_{2,\mathrm{QED}}, as the specified initial state for P1,SOA\mathcal{P}_{1,\mathrm{SOA}} and P2,SOA\mathcal{P}_{2,\mathrm{SOA}} under the one-week setting, and re-solve these two policies accordingly. We generate 1,000 independent 52-week arrival rate trajectories to evaluate the performance of the four policies. Table 12 shows the weekly average performance. Note that the weekly costs of all four policies are slightly higher than those in Table 10, because each week now absorbs carried-over workload rather than starting empty. But the superiority of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,NV\mathcal{P}_{1,\mathrm{NV}} and P2,QED\mathcal{P}_{2,\mathrm{QED}} remains similar as before, and the advantage of P2,SOA\mathcal{P}_{2,\mathrm{SOA}} over P1,SOA\mathcal{P}_{1,\mathrm{SOA}} is still marginal.

PolicyTotal cost (95% CI)Offline time (s)∑ibi\sum_i b_iAvg. ∑isi\sum_i s_iAvg. queueAvg. wait (min)Wait >10> 10 min (%)LWBS (%)
P1,NV\mathcal{P}_{1,\mathrm{NV}}1,087,236 ±\pm 11,490–15340.0014.08674.7243.683.220
P1,SOA\mathcal{P}_{1,\mathrm{SOA}}875,056 ±\pm 6,3791,69115220.002.57013.4318.360.768
P2,QED\mathcal{P}_{2,\mathrm{QED}}944,474 ±\pm 7,889–1381224.250.8024.219.280.257
P2,SOA\mathcal{P}_{2,\mathrm{SOA}}870,178 ±\pm 7,20732,022145433.702.79914.6920.820.834

Table 10. Performance comparison of four policies for the nurse-staffing case.

PolicyWeekly cost (95% CI)Weekly ∑ibi\sum_i b_iAvg. weekly ∑isi\sum_i s_iAvg. queueAvg. wait (min)Wait >10> 10 min (%)LWBS (%)
P1,NV\mathcal{P}_{1,\mathrm{NV}}1,118,401 ±\pm 73915340.0015.78283.8447.813.616
P1,SOA\mathcal{P}_{1,\mathrm{SOA}}905,817 ±\pm 44515640.002.97015.7720.720.897
P2,QED\mathcal{P}_{2,\mathrm{QED}}965,666 ±\pm 4851381246.830.9404.9910.930.304
P2,SOA\mathcal{P}_{2,\mathrm{SOA}}902,282 ±\pm 477149536.443.18016.8923.080.957

Table 12. Weekly average performance of four policies over a continuous 52-week horizon.

Conclusion

This paper proposes the SOA to solve the two-stage staffing problems with random arrival rates, aiming at more general arrival processes and service systems. Applying simulation optimization to such a problem is far from trivial and inherently accompanied by critical challenges. We effectively address these challenges by recasting the original formulation, designing an efficient coordinate-search algorithm, and constructing a two-stage staffing lookup table offline. Comprehensive numerical experiments demonstrate that the proposed SOA has significantly better staffing performance and substantially broader applicability compared to existing approaches. Promising future research directions include explicitly integrating pre-shift forecasting errors (albeit more accurate than long-term forecasts) into the simulation optimization framework, and refining the two-stage staffing lookup table to smooth out numerical anomalies and enforce strict monotonicity.

Appendix

Test of the Convergence of the Coordinate Search

For the baseline one-week case, we examine the convergence of the coordinate-search algorithm when solving the base staffing level b~~∗\tilde{\tilde{b}}^{*} in (12). We set Kout=3K_{\mathrm{out}} = 3 and record the estimated objective function value in (12) after each coordinate update. in Figure 2. It can be seen that the estimated objective function value tends to converge after the first round of outer search.

Progress of the outer coordinate search

Figure 2. Progress of the outer coordinate search. (Dashed vertical lines distinguish different outer search rounds, each covering all I=14I = 14 coordinate searches.)

Full Staffing Lookup Table for the Nurse-Staffing Case

Table 13 presents the full operational two-stage staffing lookup table of P2,SOAP_{2,\mathrm{SOA}} for the one-week (14-shift) nurse-staffing case. Note. Adjacent intervals with the same surge staffing are merged. Due to the existence of errors, the surge staffing does not strictly increase with realized arrival rates.

SunMonTueWedThuFriSat
Day shifts
Shift ii135791113
Solved base staffing bib_i62120122119116112100
Solved surge staffing s^i\hat{s}_i by λi\lambda_i interval(0,135.019]→0(0,135.019]\to0<br>(135.019,142.900]→1(135.019,142.900]\to1<br>(142.900,144.904]→2(142.900,144.904]\to2<br>(144.904,146.929]→3(144.904,146.929]\to3<br>(146.929,151.408]→4(146.929,151.408]\to4<br>(151.408,154.005]→6(151.408,154.005]\to6<br>(154.005,157.043]→7(154.005,157.043]\to7<br>(157.043,160.788]→9(157.043,160.788]\to9<br>(160.788,166.384]→11(160.788,166.384]\to11<br>(166.384,∞)→16(166.384,\infty)\to16(0,199.002]→0(0,199.002]\to0<br>(199.002,210.207]→1(199.002,210.207]\to1<br>(210.207,212.805]→2(210.207,212.805]\to2<br>(212.805,215.466]→3(212.805,215.466]\to3<br>(215.466,218.321]→5(215.466,218.321]\to5<br>(218.321,221.336]→6(218.321,221.336]\to6<br>(221.336,224.767]→7(221.336,224.767]\to7<br>(224.767,228.902]→9(224.767,228.902]\to9<br>(228.902,233.993]→12(228.902,233.993]\to12<br>(233.993,241.307]→15(233.993,241.307]\to15<br>(241.307,∞)→22(241.307,\infty)\to22(0,186.607]→0(0,186.607]\to0<br>(186.607,191.191]→1(186.607,191.191]\to1<br>(191.191,198.959]→2(191.191,198.959]\to2<br>(198.959,201.457]→3(198.959,201.457]\to3<br>(201.457,204.925]→4(201.457,204.925]\to4<br>(204.925,208.524]→6(204.925,208.524]\to6<br>(208.524,213.483]→7(208.524,213.483]\to7<br>(213.483,220.963]→11(213.483,220.963]\to11<br>(220.963,∞)→17(220.963,\infty)\to17(0,184.814]→0(0,184.814]\to0<br>(184.814,187.140]→1(184.814,187.140]\to1<br>(187.140,194.710]→2(187.140,194.710]\to2<br>(194.710,197.403]→3(194.710,197.403]\to3<br>(197.403,199.947]→4(197.403,199.947]\to4<br>(199.947,203.366]→6(199.947,203.366]\to6<br>(203.366,211.109]→8(203.366,211.109]\to8<br>(211.109,218.560]→12(211.109,218.560]\to12<br>(218.560,∞)→18(218.560,\infty)\to18(0,177.942]→0(0,177.942]\to0<br>(177.942,185.121]→1(177.942,185.121]\to1<br>(185.121,187.744]→2(185.121,187.744]\to2<br>(187.744,192.911]→3(187.744,192.911]\to3<br>(192.911,195.534]→4(192.911,195.534]\to4<br>(195.534,198.183]→5(195.534,198.183]\to5<br>(198.183,201.727]→7(198.183,201.727]\to7<br>(201.727,205.310]→8(201.727,205.310]\to8<br>(205.310,209.583]→11(205.310,209.583]\to11<br>(209.583,216.139]→13(209.583,216.139]\to13<br>(216.139,∞)→18(216.139,\infty)\to18(0,165.586]→0(0,165.586]\to0<br>(165.586,172.978]→1(165.586,172.978]\to1<br>(172.978,179.743]→2(172.978,179.743]\to2<br>(179.743,182.370]→3(179.743,182.370]\to3<br>(182.370,184.773]→4(182.370,184.773]\to4<br>(184.773,187.415]→5(184.773,187.415]\to5<br>(187.415,190.497]→6(187.415,190.497]\to6<br>(190.497,194.018]→8(190.497,194.018]\to8<br>(194.018,198.413]→10(194.018,198.413]\to10<br>(198.413,205.443]→13(198.413,205.443]\to13<br>(205.443,∞)→18(205.443,\infty)\to18(0,138.496]→0(0,138.496]\to0<br>(138.496,146.780]→1(138.496,146.780]\to1<br>(146.780,151.239]→2(146.780,151.239]\to2<br>(151.239,155.690]→3(151.239,155.690]\to3<br>(155.690,160.564]→5(155.690,160.564]\to5<br>(160.564,163.641]→6(160.564,163.641]\to6<br>(163.641,167.690]→8(163.641,167.690]\to8<br>(167.690,173.260]→10(167.690,173.260]\to10<br>(173.260,∞)→14(173.260,\infty)\to14
Night shifts
Shift ii2468101214
Solved base staffing bib_i7611310810710610390
Solved surge staffing s^i\hat{s}_i by λi\lambda_i interval(0,81.976]→0(0,81.976]\to0<br>(81.976,90.677]→1(81.976,90.677]\to1<br>(90.677,95.124]→2(90.677,95.124]\to2<br>(95.124,96.785]→3(95.124,96.785]\to3<br>(96.785,100.885]→4(96.785,100.885]\to4<br>(100.885,103.391]→5(100.885,103.391]\to5<br>(103.391,107.102]→6(103.391,107.102]\to6<br>(107.102,∞)→9(107.102,\infty)\to9(0,94.299]→0(0,94.299]\to0<br>(94.299,102.992]→1(94.299,102.992]\to1<br>(102.992,112.195]→2(102.992,112.195]\to2<br>(112.195,116.117]→4(112.195,116.117]\to4<br>(116.117,∞)→6(116.117,\infty)\to6(0,91.628]→0(0,91.628]\to0<br>(91.628,99.425]→1(91.628,99.425]\to1<br>(99.425,102.301]→2(99.425,102.301]\to2<br>(102.301,103.775]→1(102.301,103.775]\to1<br>(103.775,107.425]→2(103.775,107.425]\to2<br>(107.425,109.391]→4(107.425,109.391]\to4<br>(109.391,117.071]→6(109.391,117.071]\to6<br>(117.071,∞)→9(117.071,\infty)\to9(0,84.188]→0(0,84.188]\to0<br>(84.188,85.970]→1(84.188,85.970]\to1<br>(85.970,87.779]→2(85.970,87.779]\to2<br>(87.779,99.756]→2(87.779,99.756]\to2<br>(99.756,103.245]→3(99.756,103.245]\to3<br>(103.245,108.291]→4(103.245,108.291]\to4<br>(108.291,112.459]→8(108.291,112.459]\to8<br>(112.459,∞)→8(112.459,\infty)\to8(0,90.449]→0(0,90.449]\to0<br>(90.449,99.760]→1(90.449,99.760]\to1<br>(99.760,103.037]→2(99.760,103.037]\to2<br>(103.037,110.052]→3(103.037,110.052]\to3<br>(110.052,114.404]→4(110.052,114.404]\to4<br>(114.404,∞)→7(114.404,\infty)\to7(0,90.711]→0(0,90.711]\to0<br>(90.711,92.392]→1(90.711,92.392]\to1<br>(92.392,93.998]→0(92.392,93.998]\to0<br>(93.998,99.641]→3(93.998,99.641]\to3<br>(99.641,102.917]→5(99.641,102.917]\to5<br>(102.917,108.855]→7(102.917,108.855]\to7<br>(108.855,111.425]→8(108.855,111.425]\to8<br>(111.425,∞)→8(111.425,\infty)\to8(0,88.177]→0(0,88.177]\to0<br>(88.177,95.579]→1(88.177,95.579]\to1<br>(95.579,100.057]→2(95.579,100.057]\to2<br>(100.057,103.613]→3(100.057,103.613]\to3<br>(103.613,108.638]→4(103.613,108.638]\to4<br>(108.638,112.545]→5(108.638,112.545]\to5<br>(112.545,∞)→8(112.545,\infty)\to8

Table 13. Operational two-stage staffing lookup table of P2,SOA for the one-week (14-shift) nurse-staffing case.

References

  1. [1]Ahmed, M. A. and T. M. Alkhamis (2009). Simulation optimization for an emergency department healthcare unit in Kuwait. European Journal of Operational Research 198(3), 936–942.DOI
  2. [2]Amaran, S., N. V. Sahinidis, B. Sharda, and S. J. Bury (2016). Simulation optimization: A review of algorithms and applications. Annals of Operations Research 240(1), 351–380.
  3. [3]Bassamboo, A., R. S. Randhawa, and A. Zeevi (2010). Capacity sizing under parameter uncertainty: Safety staffing principles revisited. Management Science 56(10), 1668–1686.DOI
  4. [4]Drennan, J., A. Murphy, V. J. C. McCarthy, J. Ball, C. Duffield, R. Crouch, G. Kelly, C. Loughnane, A. Murphy, J. Hegarty, N. Brady, A. Scott, and P. Griffiths (2024). The association between nurse staffing and quality of care in emergency departments: A systematic review. International Journal of Nursing Studies 153, 104706.DOI
  5. [5]Feldman, Z., A. Mandelbaum, W. A. Massey, and W. Whitt (2008). Staffing of time-varying queues to achieve time-stable performance. Management Science 54(2), 324–338.DOI
  6. [6]Garnett, O., A. Mandelbaum, and M. Reiman (2002). Designing a call center with impatient customers. Manufacturing & Service Operations Management 4(3), 208–227.DOI
  7. [7]Green, L. V., P. J. Kolesar, and W. Whitt (2007). Coping with time-varying demand when setting staffing requirements for a service system. Production and Operations Management 16(1), 13–39.DOI
  8. [8]Hong, L. J. (2005). Discrete optimization via simulation using coordinate search. In Proceedings of the 2005 Winter Simulation Conference, pp. 803–810. IEEE.
  9. [9]Hong, L. J. and G. Jiang (2019). Offline simulation online application: A new framework of simulation-based decision making. Asia-Pacific Journal of Operational Research 36(06), 1940015.DOI
  10. [12]Kim, K. and S. Mehrotra (2015). A two-stage stochastic integer programming approach to integrated staffing and scheduling with application to nurse management. Operations Research 63(6), 1431–1451.DOI
  11. [13]Kim, S.-H. and W. Whitt (2014). Are call center and hospital arrivals well modeled by nonhomogeneous Poisson processes? Manufacturing & Service Operations Management 16(3), 464–480.DOI
  12. [16]Wang, L. and E. Demoulemeester (2023). Simulation optimization in healthcare resource planning: A literature review. IISE Transactions 55(10), 985–1007.DOI
  13. [17]Wang, X., A. Matta, N. Geng, L. Zhou, and Z. Jiang (2025). Simulation-based emergency department staffing and scheduling optimization considering part-time work shifts. European Journal of Operational Research 321(2), 631–643.DOI
  14. [18]Whitt, W. (2006). Staffing a call center with uncertain arrival rate and absenteeism. Production and Operations Management 15(1), 88–102.DOI
  15. [19]Whitt, W. and X. Zhang (2019). Forecasting arrivals and occupancy levels in an emergency department. Operations Research for Health Care 21, 1–18.DOI
  16. [20]Yousefi, M., M. Yousefi, and F. S. Fogliatto (2020). Simulation-based optimization methods applied in hospital emergency departments: A systematic review. Simulation 96(10), 791–806.DOI
  17. [21]Zhang, Y., M. L. Puterman, M. Nelson, and D. Atkins (2012). A simulation optimization approach to long-term care capacity planning. Operations Research 60(2), 249–261.DOI

Paper details

vlnc:W_h4auwxw7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents