Abstract
We consider a dynamic control problem for a GI/GI/1+GI queue with multiclass customers. The customer classes are distinguished by their interarrival time, service time, and abandonment time distributions. There is a cost c k >0 for every class k∈{1,2,…,N} customer that abandons the queue before receiving service. The objective is to minimize average cost by dynamically choosing which customer class the server should next serve each time the server becomes available (and there are waiting customers from at least two classes).
It is not possible to solve this control problem exactly, and so we formulate an approximating Brownian control problem. The Brownian control problem incorporates the entire abandonment distribution of each customer class. We solve the Brownian control problem under the assumption that the abandonment distribution for each customer class has an increasing failure rate. We then interpret the solution to the Brownian control problem as a control for the original dynamic scheduling problem. Finally, we perform a simulation study to demonstrate the effectiveness of our proposed control.
Similar content being viewed by others
Notes
Although many statements of Ito’s lemma require a twice continuously differentiable function, it follows from Problem 3.7.3 in [17] that a sufficient condition to apply Ito’s lemma is that the function is twice differentiable. See also the discussion at the end of Sect. 4.6 in [13]. This is important because the function \(\tilde{u}\) may not have a continuous second derivative, because the function g may not be continuous.
References
Akan, M., Ata, B., Olsen, T.L.: Congestion-based leadtime quotation for heterogeneous customers with convex-concave delay costs: optimality of a cost-balancing policy based on convex hull functions. Oper. Res. (2011, in press)
Armony, M., Ward, A.R.: Fair dynamic routing in large-scale heterogeneous-server systems. Oper. Res. 58(3), 624–637 (2010)
Ata, B., Olsen, T.L.: Near-optimal dynamic leadtime quotation and scheduling under convex-concave customer delay costs. Oper. Res. 57(3), 753–768 (2009)
Ata, B., Olsen, T.L.: Congestion-based leadtime quotation and pricing for revenue maximization with heterogeneous customers. Working Paper, Northwestern University, Evanston, IL (2011)
Ata, B., Rubino, M.: Dynamic control of a make-to-order parallel-server system with cancellations. Oper. Res. 57(1), 94–108 (2009)
Ata, B., Tongarlak, M.H.: On scheduling a multiclass queue with abandonments under general delay costs. Working paper (2012)
Atar, R., Giat, C., Shimkin, N.: The \(\frac{c\mu}{\theta}\) rule for many-server queues with abandonment. Oper. Res. 58(5), 1427–1439 (2010)
Atar, R., Giat, C., Shimkin, N.: On the asymptotic optimality of the \(\frac{c\mu}{\theta}\) rule under ergodic cost. Queueing Syst. 67(2), 127–144 (2011)
Cox, D., Smith, W.: Queues. Methuen, London (1961)
Down, D.G., Koole, G., Lewis, M.: Dynamic control of a single server system with abandonments. Queueing Syst. 67(1), 63–90 (2011)
Ghosh, A., Weerasinghe, A.: Optimal buffer size for a stochastic processing network in heavy traffic. Queueing Syst. 55, 147–159 (2007)
Harrison, J.M.: Brownian Motion and Stochastic Flow Systems, 1st edn. Wiley Series in Probability and Statistics. Wiley, New York (1985)
Harrison, J.M.: Brownian models of open processing networks: canonical representation of workload. Ann. Appl. Probab. 10(1), 75–103 (2001)
Harrison, J.M., Van Mieghem, J.A.: Dynamic control of Brownian networks: state space collapse and equivalent workload formulations. Ann. Appl. Probab. 7(3), 747–771 (1997)
Harrison, J.M., Zeevi, A.: Dynamic scheduling of a multiclass queue in the Halfin–Whitt heavy traffic regime. Oper. Res. 51, 243–257 (2004)
Hartman, P.: Ordinary Differential Equations. Wiley, New York (1964)
Karatzas, I., Shreve, S.E.: Brownian Motion and Stochastic Calculus, 2nd edn. Graduate Texts in Mathematics, vol. 113. Springer, New York (1988)
Lee, C., Weerasinghe, A.: Convergence of a queueing system in heavy traffic with general patience-time distributions. In: Stochastic Processes and their Applications, vol. 121, pp. 2507–2552 (2011)
Mandelbaum, A., Momcilovic, P.: Queues with many servers and impatient customers. Math. Oper. Res. 37(1), 41–65 (2012)
Mandelbaum, A., Stolyar, A.L.: Scheduling flexible servers with convex delay costs: heavy-traffic optimality of the generalized cμ-rule. Oper. Res. 52(6), 836–855 (2004)
Mandelbaum, A., Zeltyn, S.: Call centers with impatient customers: many-serve asymptotics of the M/M/N+G queue. Queueing Syst. 51(3/4), 361–402 (2005)
Van Mieghem, J.: Dynamic scheduling with convex delay costs: the generalized cμ rule. Ann. Appl. Probab. 5(3), 809–833 (1995)
Ramanan, K., Kang, W.: Fluid limits of many-server queues with reneging. Ann. Appl. Probab. 20, 2204–2260 (2010)
Reed, J.E., Ward, A.R.: Approximating the GI/GI/1+GI queue with a nonlinear drift diffusion: hazard rate scaling in heavy traffic. Math. Oper. Res. 33(3), 606–644 (2008)
Reiman, M.I.: The heavy traffic diffusion approximation for sojourn times in Jackson networks. In: Disney, R.L., Ott, T.J. (eds.) Applied Probability-Computer Science, the Interface. Birkhauser, Boston (1982)
Sigman, K.: Lecture Notes on Stochastic Modeling I (2011). Available from Karl Sigman’s web site
Smith, W.: Various optimizers for single-stage production. Nav. Res. Logist. Q. 3, 59–66 (1956)
Tezcan, T., Dai, J.G.: Optimal control of parallel server systems with many servers in heavy traffic. Queueing Syst. 59, 95–134 (2008)
Tezcan, T., Dai, J.G.: Dynamic control of N-systems with many servers: asymptotic optimality of a static priority policy in heavy traffic. Oper. Res. 58(1), 94–110 (2010)
Ward, A.R., Glynn, P.W.: A diffusion approximation for a Markovian queue with reneging. Queueing Syst. 43(1/2), 103–128 (2003)
Ward, A.R., Glynn, P.W.: A diffusion approximation for a GI/GI/1 queue with balking or reneging. Queueing Syst. 50(4), 371–400 (2005)
Weber, T.: Optimal Control Theory with Applications in Economics. MIT press, Cambridge (2011)
Whitt, W.: Stochastic Process Limits. Springer, New York (2002)
Acknowledgements
We would like to thank Baris Ata, Ramandeep Randhawa, and Ananda Weerasinghe for helpful discussions regarding how to solve the HJB equations.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
The proofs of all propositions and lemmas in the paper appear in this Appendix, in the order that they appear in the main body of the paper.
1.1 Proof of Proposition 4.1
The first step is to define the reduced Brownian control problem, which has the one-dimensional state-descriptor \(\hat{W}\). The reduced Brownian control problem is to
using a control process \((\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat {Q}_{N},\hat{I} )\) such that
The control \((\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat{Q}_{N},\hat {I} )\) is admissible if it is \(\mathcal{F}_{t}\)-adapted and satisfies (36). The control \((\hat {Q}_{1}^{\star},\hat{Q}_{2}^{\star},\ldots,\hat{Q}_{N}^{\star},\hat {I}^{\star} )\) is optimal if it is admissible and for any admissible \((\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat{Q}_{N},\hat {I} )\)
where
The following two claims, which are Propositions 2 and 3 in [5] adapted to our setting, complete the proof. The claims are stated without proof because the arguments in [5] do not rely on their assumption of linear drift, and so apply to our nonlinear drift setting, also.
Claim 1
(Proposition 2 in [5])
The Brownian control problem (6)–(8) is equivalent to the reduced Brownian control problem (35)–(36) in the following sense:
-
Every admissible control \((\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat {Q}_{N},\hat{I} )\) for the reduced Brownian control problem yields an admissible control \(\hat{Y}\) for the Brownian control problem having \(\hat{\mathcal{C}}(\hat{Y}) = \hat{\mathcal{C}}_{R} (\hat{Q}_{1},\hat {Q}_{2},\ldots,\hat{Q}_{N},\hat{I} )\).
-
Similarly, for any admissible control \(\hat{Y}\) for the Brownian control problem, there exists an admissible control \((\hat {Q}_{1},\hat{Q}_{2},\ldots,\hat{Q}_{N},\hat{I} )\) for the reduced Brownian control problem having \(\hat{\mathcal{C}}(\hat{Y}) \geq\hat {\mathcal{C}}_{R} (\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat{Q}_{N},\hat {I} )\).
Claim 2
(Proposition 3 in [5]) The reduced Brownian control problem (35)–(36) is equivalent to the workload control problem in the following sense:
-
Every admissible control \((q,\hat{I})\) for the workload control problem yields an admissible control \((\hat{Q}_{1},\hat {Q}_{2},\ldots,\hat{Q}_{N},\hat{I} )\) for the reduced Brownian control problem having \(\hat{\mathcal{C}}_{R} (\hat{Q}_{1},\hat {Q}_{2},\ldots,\hat{Q}_{N},\hat{I} ) = \hat{\mathcal{C}}_{w}(q,\hat{I})\).
-
Similarly, for any admissible control \((\hat{Q}_{1},\hat {Q}_{2},\ldots,\hat{Q}_{N},\hat{I} )\) for the reduced Brownian control problem, there exists an admissible control for the workload control problem having \(\hat{\mathcal{C}}_{w}(q, \hat{I}) \leq\hat {\mathcal{C}}_{R} (\hat{Q}_{1},\hat{Q}_{2},\ldots,\hat{Q}_{N},\hat {I} )\). □
1.2 Proof of Corollary 4.1
It follows from Proposition 4.1 that given the control \(\hat{Y}^{\star}\) that is optimal for the Brownian control problem (6)–(8), then there exists an admissible control \((q,\hat{I})\) for the workload control problem having
Suppose that \((q,\hat{I})\) is not optimal for the workload control problem. Then there exists an admissible control \((q^{\star}, \hat {I}^{\star})\) where
It follows from Proposition 4.1 that there exists an admissible control \(\hat{Y}\) for the Brownian control problem having
This is a contradiction, because (37)–(39) implies \(\hat{\mathcal{C}}(\hat{Y}) < \hat{\mathcal{C}}(\hat{Y}^{\star})\). □
1.3 Proof of Proposition 4.2
The following notation is useful in the proof. For any w≥0, let
Then, since \(\hat{W}^{\star}(0) = \hat{w}\), for the second statement of the proposition, we must show that \(E_{\hat{w}} [\hat{W}^{\star}(t) ]/t \rightarrow0\) as t→∞. Note that we explicitly subscript the expectation because it is useful in the proof to also consider what happens when the initial state differs from \(\hat{w}\).
We observe that the process \(\hat{W}^{\star}\) is a delayed regenerative process with regeneration point 0. The initial cycle length is \(T_{0} := \inf \{ t \geq0: \hat{W}^{\star}(t) = 0 \}\). The remaining i.i.d. cycles are defined as the process starting from 0, reaching \(\overline{w}>0\), and returning to 0, and we let τ represent the length of one of those cycles. Suppose we can show that
-
(i)
\(\hat{W}^{\star}\) is positive recurrent; that is, that \(E_{\hat {w}}[\tau] < \infty\);
-
(ii)
\(E_{\hat{w}} [ \int_{T_{0}}^{T_{0} + \tau} \hat{W}^{\star}(s)\,ds ] < \infty\);
-
(iii)
\(E_{\hat{w}} [ \int_{0}^{T_{0}} \hat{W}^{\star}(s)\,ds ] < \infty\).
Then, also noting that \(\int_{0}^{T_{0}} \hat{W}^{\star}(s)\,ds < \infty\) almost surely, it follows from the renewal reward theorem for regenerative processes (see, for example, Theorem 2.3 in Sect. 13 on Regenerative Processes of [26]) that
In the next paragraph, we establish that (41) implies
To show (42), it is enough to show that
To show (43), the first step is to recall that \(\hat{\xi }(\cdot)\) in (15) has the same distribution as \(\theta \cdot+ \sigma\hat{B}(\cdot)\) for \(\hat{B}\) a standard Brownian motion, and then apply Ito’s lemma to find
Since any sample path of \(\hat{W}^{\star}(t)^{2}\) is continuous almost surely, for any t>0, \(\int_{0}^{t} \hat{W}^{\star}(s)^{2} \,ds < \infty\) almost surely and \(\hat{M}(t) := \int_{0}^{t} \hat{W}^{\star}(s) \,d\hat{B}(s)\) is a local martingale. Then there exists a sequence of stopping times {τ k ,k=1,2,…} such that τ k <τ k+1 for all k=1,2,…,τ k →∞ as k→∞ almost surely, and \(\hat{M}(t \wedge\tau_{k})\) is a martingale for each k=1,2,… , so that \(E [ \hat{M}(t \wedge\tau_{k}) ] = 0\). Then
where
-
(45) follows from the definition of the sequence {τ k ,k=1,2,…};
-
(46) follows from Fatou’s lemma;
-
(47) follows from (44) and the fact that \(\hat{M}(t \wedge\tau_{k})\) is a martingale;
-
(48) follows from the monotone convergence theorem.
Finally, it follows from (45)–(48) that to establish (43), it is enough to establish (41).
In summary, the proof is complete once we show (i)–(iii). For this, it is helpful to define the operator
where
and u is any twice differentiable function having domain ℜ+. In the following, assuming \(\overline{w} > \hat{w}\), we show that for
for any k∈{0,1,2,…}
and
This is sufficient to complete the proof because when \(\hat{w} = 0\), (49) and (50) imply (i) and (ii) by letting k=0 and k=1, and for \(\hat{w} >0\),
which implies (iii) by letting k=1.
The argument to show (49)
For any x≥0, let
Then it is straightforward to check that
solves
Moreover, since r(x)<0 for x>0, u(x) is decreasing. This implies that u(x)≥0 for \(x\in [0,\overline{w} ]\) because \(u (\overline{w} )=0\). Furthermore, there exists a twice differentiable function \(\tilde{u}\) having domain ℜ+ such that \(\tilde{u}(x)=u(x)\) for \(x\in [0,\overline{w} ]\) and \(\tilde{u}'\) is bounded.
From Ito’s lemmaFootnote 1, by recalling \(\hat{\xi}(\cdot) \stackrel{D}{=} \theta\cdot+ \sigma\hat {B}(\cdot)\) for a standard Brownian motion \(\hat{B}(\cdot)\) with \(\hat{B}(0)=0\),
Since \(\tilde{u}'\) is bounded, the stochastic integral in the above is a martingale and, therefore, so is
Then
is also a martingale, and so
Since \(\hat{W}^{\star} (t\wedge T^{\overline{w}} )\in [0,\overline{w}]\), \(\tilde{u} (\hat{W}^{\star} (t\wedge T^{\overline{w}} ) )\geq0\), and so
Taking the limit as t→∞ in the above and using the monotone convergence theorem implies
which establishes (49).
The argument to show (50)
For any \(\tilde{w} > \overline{w}\), let
Next, for any x≥0, let
where
Then it is straightforward to check that
solves
Note that there can be at most one x q >0 such that q(x)>0 for all x∈[0,x q ) and q(x)<0 for all x>x q . Then \(v_{\tilde {w}}(0)=v_{\tilde{w}} (\tilde{w} )=0\) implies that such x q exists and \(x_{q}\in (0,\tilde{w} )\) so that \(v_{\tilde{w}}(x)\) is strictly increasing for x∈[0,x q ] and strictly decreasing otherwise. This implies that \(v_{\tilde{w}}(x) \geq0\) for all \(x\in [0,\tilde{w} ]\). Furthermore, there exists a twice differentiable function \(\tilde{v}_{\tilde{w}}\) having domain ℜ+ such that \(\tilde{v}_{\tilde{w}}(x)=v_{\tilde{w}}(x)\) for all \(x\in [0,\tilde{w} ]\) and \(\tilde{v}_{\tilde{w}}'\) is bounded.
From Ito’s lemma, by recalling \(\hat{\xi}(\cdot) \stackrel{D}{=} \theta \cdot+ \sigma\hat{B}(\cdot)\) for a standard Brownian motion \(\hat {B}(\cdot)\) with \(\hat{B}(0)=0\),
Since \(\tilde{v}'\) is bounded, the stochastic integral in the above is a martingale and, therefore, so is
as well as
If we renew \(\hat{W}^{\star}\) once the stochastic process reaches \(\overline{w}>\hat{w}\), we have \(\hat{W}^{\star}(s) \in(0,\tilde{w})\) for all \(s \in[0, t \wedge T_{0}^{\tilde{w}})\). Then, for all \(s \in[0, t \wedge T_{0}^{\tilde{w}})\), \(d\hat{I}^{\star}(s) = 0\). Therefore, since \(\hat{I}^{\star}\) is continuous, it follows that
Then the same argument as in the previous paragraph shows that
The monotone convergence theorem then implies
Finally, to complete the proof, it follows from (51) and (52) that it is enough to show there exists M>0 that does not depend on \(\tilde{w}\) such that
To prove (53), observe that, for any \(\tilde {w} > \overline{w}\), we have
Since g(x)→−∞ as x→∞, we have
Also, we argue that
To prove, observe that there exists y ⋆>0 such that g(x)≤0 for all x≥y ⋆ because g(x)→−∞ as x→∞. Then for any \(\tilde{w} \geq y^{\star}\), we have
implying (54). Then, we use L’Hopital’s rule to conclude
Since it is straightforward to see that
the proof is completed. □
1.4 Proof of Lemma 4.1
It is enough to show that the function ϕ(x,w) is jointly continuous in (x,w) and Lipschitz continuous in w. This is because the fact that there exists a unique solution w κ of IVP(κ) on a compact interval of ℜ+ then follows from Picard’s existence theorem (see, for example, Proposition 2.4 of [32]). The fact that there exists a unique solution w κ of IVP(κ) on ℜ+ follows an iteration argument, which we omit because it is standard.
To see that ϕ(x,w) is continuous in (x,w), let ϵ>0 be arbitrarily small. We show that, for i∈{1,2}, if x i ≥0 and w i ≥0 are such that
for
then
Define
so that
Next, define
where
is such that
so that
Note that such a \(p \in\mathcal{P}\) can be found because for any \(p \in\mathcal{P}\), \(\tilde{q}_{1}^{2} \geq0\), and, for fixed x 1 and x 2, \(\tilde{q}_{2}^{2}\) is a continuous function of p having \(\tilde {q}_{1}^{2}(p) \rightarrow\mu_{2} x_{2} > 0\) as \(p \rightarrow q_{1}^{1\star} / (\mu_{1}(x_{1} - x_{2}))\). Without loss of generality, assume (x 1,w 1) and (x 2,w 2) are such that
Then
Since
and
where the last inequality follows from the definition of \((\tilde {q}_{1}^{2},\tilde{q}_{2}^{2},\ldots,\tilde{q}_{N}^{2} )\), it follows that
so that
To see that ϕ(x,w) is Lipschitz continuous in w, note that it follows from (55) that when x 2=x 1=x≥0,
1.5 Proof of Proposition 4.3
From Theorem 4.1, it is sufficient for the proof to construct v:ℜ+→ℜ that is twice continuously differentiable and κ>0 that solves the Bellman equation (20)–(21). Define
for
Then, v′ solves
The first step in showing that (v,κ ⋆) solves the Bellman equation is to show that (20) is satisfied. Suppose we can show that v′(x)∈[0,c 1 μ 1] for all x≥0. Then, we have
for any x≥0. To see this, observe that
for any \(q \in\mathcal{A}(x)\). Next, note that as a function of q 1, \(q_{1}^{2} / (\mu_{1}f_{1} )+ (\mu_{2} (x-q_{1}/\mu_{1} ) )^{3}/ (\mu_{2}f_{2}^{2} )\) is convex because
Therefore, \(\mbox{argmin}_{q_{1}\in[0,\mu_{1}x]} \{ q_{1}^{2} / (\mu_{1}f_{1} )+ (\mu_{2} (x-q_{1}/\mu_{1} ) )^{3}/ (\mu_{2}f_{2}^{2} ) \}\), for any given x≥0, is derived by solving
Since the solution of (57) is \(q_{1}^{\star}(x)\) defined in Sect. 4.2.1, we have
which implies by (56) that (20) is satisfied.
The argument that v′(x)∈[0,c 1 μ 1] for all x≥0 is by contradiction; suppose not. Then, since v′(0)=0 and v′(x)→c 1 μ 1 as x→∞ follows from L’Hopital’s rule, there must exist x 1<x 2 such that v′(x 1)=v′(x 2)>c 1 μ 1 and v″(x 1)>0, v″(x 2)<0. From (56),
Subtracting the bottom equation from the top yields
This is a contradiction because v″(x 1)−v″(x 2)>0, c−v′(x 1)<0, and g(x 1)−g(x 2)<0 since g is a strictly increasing function.
The second step in showing that (v,κ ⋆) solves the Bellman equation is to show that (21) is satisfied. It is immediate that v′(0)=0. The fact that v′ is bounded follows from the fact that v′ is continuous and v′(x)→c 1 μ 1 as x→∞ by L’Hopital’s rule. To see that v′ is nonnegative, note that it follows from its definition that either v′(x)≥0 for all x, in which case the proof is complete, or there exists x 0 such that v′(x)>0 for all x<x 0 and v′(x)<0 for all x>x 0. This is a contradiction because v′(x)→c 1 μ 1>0 as x→∞, and so there cannot exist such an x 0. □
1.6 Proof of Lemma 4.2
Note that c 1 μ 1 γ 1>c 2 μ 2 γ 2 and c 2 μ 2>c 1 μ 1 implies that γ 1>γ 2,
and
Then
Since κ 1(L) and κ 2(L) are continuous in L, we conclude that there exists L>0 such that κ 1(L)=κ 2(L). Since κ 2(L)>0 for all L≥0, it must be that at L such that κ 1(L)=κ 2(L), κ 1(L)=κ 2(L)>0. □It is useful for the proof of Proposition 4.4 to note that
for all x∈ℜ (see, for example, Lemma EC.3 in [2] to see that xΦ(x)+ϕ(x)>0 for all x∈ℜ), so that Ψ is a strictly decreasing function. Also, Ψ(x)→∞ as x→−∞ and Ψ(x)→0 as x→∞.
1.7 Proof of Proposition 4.4
From Theorem 4.1, it is sufficient for the proof to construct v:ℜ+→ℜ that is twice continuously differentiable and κ>0 that solves the Bellman equation (20)–(21). We first do this when c 1 μ 1 γ 1>c 2 μ 2 γ 2 and c 1 μ 1≥c 2 μ 2, so that a static priority control is optimal, and second do this when c 1 μ 1 γ 1>c 2 μ 2 γ 2 and c 1 μ 1<c 2 μ 2, so that a dynamic priority control is optimal.
The proof when c 1 μ 1 γ 1>c 2 μ 2 γ 2 and c 1 μ 1≥c 2 μ 2. Define
for
Note that κ ⋆>0 because
and \(\varPsi (-\sqrt{\frac{2}{\sigma^{2} \gamma_{2}}}\theta )>0\). Also,
so that it is straightforward to see v′(0)=0 and v′(x) increases to c 2 μ 2 as x→∞. Furthermore, v′ solves
To show that (v,κ ⋆) solves the Bellman equations (20)–(21), first note that it is immediate from the representation of v′ in (58) and the properties of Ψ that (21) is satisfied. To see that (20) is also satisfied, we show that
Then it follows from (59) and (60) that (20) holds, i.e., that
We show that (60) holds by considering each of the two possibilities: c 1 μ 1=c 2 μ 2 or c 1 μ 1>c 2 μ 2. In the case that c 1 μ 1=c 2 μ 2, it follows from the fact that c 1 μ 1 γ 1>c 2 μ 2 γ 2 that γ 1>γ 2, so that
and (60) holds. Otherwise, in the case that c 1 μ 1>c 2 μ 2, first note that to show (60) holds it is equivalent to show
because when (61) holds
Then, in the case that c 1 μ 1>c 2 μ 2 and γ 1<γ 2, (61) holds because (61) is equivalent to
which holds because (c 1 μ 1 γ 1−c 2 μ 2 γ 2)/(γ 1−γ 2)<0 and v′(x)≥0 for all x≥0 from its explicit expression in (58). Finally, in the case that c 1 μ 1>c 2 μ 2 and γ 1≥γ 2, (61) holds because (61) is equivalent to
which holds because v′(x)≤c 2 μ 2 for all x≥0 by (58), c 2 μ 2<c 1 μ 1 by assumption, and
The proof when c 1 μ 1 γ 1>c 2 μ 2 γ 2 and c 1 μ 1<c 2 μ 2 is true. Define
and
Then, it is straightforward to verify that \(v_{1}'\) and \(v_{2}'\) in (62) and (63) solve
with initial conditions
and that
It follows from (64) and (65) that \(v_{1}'' ( L^{\star}) = v_{2}'' ( L^{\star})\). Therefore,
is twice-continuously differentiable.
The first step in showing (v,κ ⋆) solves the Bellman equation is to show that (20) is satisfied. For this, observe that
Therefore, if we can show that
and
it follows from (64) that (20) holds, i.e., that
Furthermore, it also follows that
To see (67) holds, replace κ ⋆ in (63) by κ 2(L ⋆) in Lemma 4.2 so that
Then \(v_{2}'(x)\) for all x≥L ⋆ is strictly increasing function in x because −Ψ(x) is strictly increasing in x for all x∈ℜ. So (67) holds since
The argument that (66) holds is by contradiction. Suppose that there exists x 0<L ⋆ for which \(v_{1}'(x_{0})> (c_{1}\mu_{1}\gamma_{1}-c_{2}\mu_{2}\gamma_{2} )/ (\gamma_{1}-\gamma_{2} )\). Then there exists x 1∈(x 0,L ⋆) such that \(v''_{1}(x_{0})<0\) and \(v'_{1}(x_{1})= (c_{1}\mu_{1}\gamma_{1}-c_{2}\mu_{2}\gamma_{2} )/ (\gamma_{1}-\gamma_{2} )\). From (64),
so that, subtracting the bottom equation from the top shows that
This is a contradiction because \(v_{1}'' ( L^{\star}) = v_{2}'' ( L^{\star})\geq0\) (since \(v_{2}'\) is a strictly increasing function), \(v_{1}''(x_{1}) < 0\), and x 1<L ⋆ implies that
The second step in showing (v,κ ⋆) solves the Bellman equation is to show that (21) is satisfied, i.e., that
-
(i)
v′(0)=0;
-
(ii)
v′ is bounded;
-
(iii)
v′ is nonnegative.
For (i), recall that \(v'_{1}(0)=0\). For (ii), we can show that \(v_{2}'(x) \rightarrow c_{1} \mu_{1}\) as x→∞ by noting that Ψ(x)→0 as x→∞ in (68). For (iii), by calculus and algebra, it can be shown that \(v_{1}'\) in (62) has the equivalent representation
Then, depending on how large κ ⋆ is, \(v'_{1}\) is either nonnegative for all x≥0 or there exists a unique x 0>0 such that \(v_{1}'(x) > 0\) for all x<x 0 and \(v_{1}'(x) < 0\) for all x>x 0. Since \(v'_{1}(L^{\star}) = v'_{2}(L^{\star})= \frac{c_{1} \mu_{1} \gamma_{1} - c_{2} \mu_{2} \gamma_{2}}{\gamma_{1} - \gamma_{2}}>0\), x 0>L ⋆ if such x 0 exists. Therefore, \(v'_{1}\) is nonnegative for all x<L ⋆. Note that \(v'_{2}(x) >0\) for all x>L ⋆ because \(v_{2}'(L^{\star}) >0\) and \(v_{2}'\) is strictly increasing. Therefore, v′ is nonnegative. □
1.8 Proof of Lemma 6.1
We already proved that ϕ(x,w) is jointly continuous in (x,w) in the proof of Lemma 4.1. Therefore, we can apply Theorem 2.1 in [16] to conclude that w κ (x) is jointly continuous in (x,κ). □
1.9 Proof of Lemma 6.2
Let w 1>w 2>0. Then, recalling that θ≤0,
1.10 Proof of Lemma 6.3
The proof for Lemma 6.3 is identical to the proof of Lemma 4 in the Appendix of [5], and so is omitted. □
1.11 Proof of Lemma 6.4
Since ϕ(0,0)=0, it follows from the definition of IVP(κ) that
Hence, the fact that \(w'_{\kappa}\) is continuous on ℜ+ and the existence of x κ such that \(w'_{\kappa} (x_{\kappa } )<0\) imply that the set
is not empty and
To complete the proof, we have to show two things; (i) \(\overline {\mathcal{E}}:= \{ x>x_{\kappa}:w'_{\kappa}(x)=0 \} \) is empty and (ii) \(w'_{\kappa}(x)\geq0\) for all x∈[0,x 0].
Proof of (i)
The argument that \(\overline{\mathcal{E}}\) is empty is by contradiction; suppose not that \(\overline{\mathcal{E}}\) is not empty. We can define
Then, by the definitions of x 0 and x κ, we have
and
which from the definition of IVP(κ) implies that
Assume we can show that
Also, (69) and Lemma 6.2 show that
which contradicts (70).
To establish (71), assume we can show that
Let
so that
Since x κ>x 0, there exists \(q^{0}\in\mathcal{A} (x_{0} )\) such that \(q_{i}^{0}\leq q_{i}^{\kappa\star}\) for all i∈{1,2,…,N} with at least one of the inequalities being strict. This implies that \(m_{i} (q_{i}^{0} )\leq m_{i} (q_{i}^{\kappa\star} )\) for all i∈{1,2,…,N}, because h i is increasing. Since (73) implies that (c i −w κ (x 0)/μ i )>0 for all i∈{1,2,…,N}, it follows that
The definition of ϕ implies that
Together, (74) and (75) show (71).
We now establish (73). First suppose that w κ (x 0)≥max(c 1 μ 1,c 2 μ 2,…,c N μ N ). Then c i −w κ (x 0)/μ i ≤0 for all i∈{1,2,…,N}, so that for any \(q\in\mathcal{A} (x_{0} )\)
which implies ϕ(x 0,w κ (x 0))≤0. This is a contradiction because it follows from IVP(κ) that
but we have assumed κ>0. Next, let c i μ i for i∈{1,2,…,N} be ordered so that \(c_{i_{1}}\mu_{i_{1}}\leq c_{i_{2}}\mu_{i_{2}}\leq \cdots \leq c_{i_{N}}\mu_{i_{N}}\). We may assume that at least one of inequalities is strict because otherwise we have min(c 1 μ 1,c 2 μ 2,…,c N μ N )=max(c 1 μ 1,c 2 μ 2,…,c N μ N ) and the proof is complete. Let s∈{1,2,…,N−1} be the largest value with \(c_{i_{s}}\mu_{i_{s}}<c_{i_{s+1}}\mu_{i_{s+1}}\) and suppose \(c_{i_{s}}\mu_{i_{s}}\leq w_{\kappa} (x_{0} )<c_{i_{s+1}}\mu_{i_{s+1}}\). Then, \(c_{i_{j}}\mu_{i_{j}}-w_{\kappa} (x_{0} )>0\) for all j∈{s+1,s+2,…,N} and \(c_{i_{j}}\mu_{i_{j}}-w_{\kappa} (x_{0} )\leq0\) for all j∈{1,2,…,s} so that
where \(q^{s}\in\mathcal{A} (x_{0} )\) such that \(q_{i_{j}}^{s}=0\) for all j∈{s+1,s+2,…,N}. This is again a contradiction for the same reason as earlier in this paragraph. If we repeat this argument, we can prove that w κ (x 0)<min(c 1 μ 1,c 2 μ 2,…,c N μ N ) and this completes the proof of (i).
Proof of (ii)
Finally, to complete the proof of the lemma, it remains to prove that \(w'_{\kappa}(x)\geq0\) for all x∈[0,x 0]. Suppose there exists x 1∈(0,x 0) such that \(w'_{\kappa} (x_{1} )<0\). Then, by applying the same argument in the proof of (i) to x 1, we conclude that \(w'_{\kappa}(x)<0\) for all x>x 1. This is a contradiction because \(w'_{\kappa} (x_{0} )=0\) and x 0>x 1. □
1.12 Proof of Lemma 6.5
If \(\kappa\in\mathcal{D}\), then by definition of \(\mathcal{D}\), there exists x κ >0 such that w κ (x) is decreasing if x>x κ . There are two possibilities: (i) w κ (x) is bounded for all x≥0, or (ii) w κ (x)→−∞ as x→∞. We show that case (i) leads to a contradiction, which is enough to complete the proof.
Suppose w κ (x) is bounded for x≥0. Then lim x→∞ w κ (x) exists and is finite, which implies \(w_{\kappa}'(x) \rightarrow0\) as x→∞, and so, from the definition of IVP(κ),
Next, since \(\kappa\in\mathcal{D}\), there exists \(\overline{w}\) such that \(w_{\kappa}(x) < \overline{w}\) for all x>0. Then, by Lemma 6.2,
Let x>x κ and define
so that
It follows from (73) in the proof of Lemma 6.4 that for x 0 such that \(w_{\kappa}'(x_{0}) = 0\), w κ (x 0)<min(c 1 μ 1,c 2 μ 2,…,c N μ N ), so that we may assume \(\overline{w} < \min(c_{1},\mu_{1}, c_{2} \mu_{2},\ldots,c_{N}\mu_{N})\). Then \(c_{k}\mu_{k} - \overline{w} >0\) for all k∈{1,2,…,N}, and it follows from (78) that
because \(\max (q_{1}^{\star} (x ),q_{2}^{\star} (x ),\ldots,q_{N}^{\star} (x ) )\rightarrow\infty\) as x→∞ from the fact that \(\sum_{k=1}^{N}q_{k}^{\star } (x )/\mu_{k}=x\) for every x>0. Finally, (79) and (77) imply that
which contradicts (76). □
1.13 Proof of Lemma 6.6
Fix κ>0. First note that if there exists x 0>0 such that w κ (x 0)≥min(c 1 μ 1,c 2 μ 2,…,c N μ N ), then by the same argument as the one in the second to the last paragraph of the proof of Lemma 6.4,
It follows from (80) that
Since w κ is continuous, and the same argument as above shows that \(w_{\kappa}'(x) >0\) whenever w κ (x)≥min(c 1 μ 1,c 2 μ 2,…,c N μ N ), it follows that \(w_{\kappa}'(x) >0\) for all x≥x 0.
Next, we claim that there can exist at most one \(\underline{x} >0\) such that \(w_{\kappa}(\underline{x}) = \min (c_{1}\mu_{1},c_{2}\mu_{2},\ldots,c_{N}\mu_{N} )\). To see this, note that the argument in the previous paragraph shows that \(w_{\kappa}'(x) >0\) for all \(x \geq \underline{x}\).
Finally, to complete the proof, we must show that w κ (x) is increasing on \([0,\underline{x})\). This argument is by contradiction. Assume there exists x 1 such that \(w_{\kappa}'(x_{1}) < 0\). Then, by Lemma 6.4, \(\kappa\in\mathcal{D}\), and there exists x κ such that w κ is decreasing on [x κ ,∞). This is a contradiction, because the previous two paragraphs show that \(w_{\kappa}'(x) >0\) for all \(x \geq\underline{x}\). □
Rights and permissions
About this article
Cite this article
Kim, J., Ward, A.R. Dynamic scheduling of a GI/GI/1+GI queue with multiple customer classes. Queueing Syst 75, 339–384 (2013). https://doi.org/10.1007/s11134-012-9325-7
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-012-9325-7
Keywords
- Scheduling control in multiclass queue
- Customer abandonment
- Customer reneging
- Customer impatience
- Diffusion approximation
- Brownian control problem