2.6 Structural properties of the integrated system
2.6.2 Effect of pooling demand, inventories and service capacity
Definition 2.6.1. We say a subset{φ1, . . . , φN} ⊆ {1, . . . , J}of locations is homogeneous if
λφk =λφ` and γφk =γφ`, ∀k, `∈ {1, . . . , N}.
A homogeneous set{φ1, . . . , φN}of locations is1-homogeneous ifγφk = 1, ∀k∈ {1, . . . , N}
holds.
Next we show that pooling of homogeneous systems (or locations) reduces the optimal base stock levels and the costs. This is done first for 1-homogeneous systems. We will verify this twofold. First by proving the following proposition.
Proposition 2.6.2. Consider an ergodic network as in Section 2.2 that includes (among others) N locations with demand streams of intensities bλφ1 = · · · = λbφN, with optimal base stock levelsbb∗φ
1, . . . ,bb∗φ
N and corresponding costsbgφ1(bb∗φ
1), . . . ,bgφN(bb∗φ
N). Assume that these locations are1-homogeneous, i.e. bγφi = (νpbφi)/bλφi = 1, i= 1, . . . , N. When the N arrival streams are pooled to arrive at a single location, denoted by φ, with demand rate bλφ1+· · ·+λbφN =λbφ1 ·N =:λφ andpφ:=pbφ1 ·N, then for the optimal base stock level b∗φ and costs gφ(b∗φ) at the pooled location φ the following holds:
bb∗φ1 +· · ·+bb∗φN (≈√
N·b∗φ> b∗φ if bb∗φ1 >1,
&√
N·b∗φ> b∗φ if bb∗φ
1 = 1 and
gbφ1(bb∗φ1) +· · ·+bgφN(bb∗φN) (≈√
N·gφ(b∗φ) if bb∗φ
1 >1
&N ·gφ(b∗φ) if bb∗φ
1 = 1 )
> gφ(b∗φ).
An explanation for this decrease in case of pooling is probably that the system with more stations generates more variability in its performance metrics.
This proposition shows that the optimal base stock level and associated costs in the pooled system are smaller than the sum of the individual components in the non-pooled system.
Proof of Proposition 2.6.2. For technical reasons we investigate the reversed process of pooling: Splitting demand and inventory. For this we distribute the demand of rate λφ at locationφtoN locations{φ1, . . . , φN}with reduced demand of ratebλφk :=λφ/N and pbφk :=pφ/N, ∀k∈ {1, . . . , N}.
We first collect necessary prerequisites and remark that γφ = 1 implies bγφk = 1, ∀k ∈ {1, . . . , N} (and vice versa), i.e. the property of 1-homogeneity is hereditary for the N split locations.
Forγφ= 1 the cost function is
gφ(bφ) = (cs,φ+ch,J+1)·bφ+cls,φ·λφ·P(Yφ= 0) + (ch,φ−ch,J+1)·E(Yφ)
(2.5.2)+(2.5.3)
= (cs,φ+ch,J+1)·bφ+cls,φ·λφ· 1
(bφ+ 1)+ (ch,φ−ch,J+1)·bφ 2
=
cs,φ+1
2(ch,φ+ch,J+1)
·bφ+cls,φ·λφ· 1 (bφ+ 1).
2.6. Structural properties of the integrated system
To simplify calculations we will analyse optimal points b∗φ ∈ [1,∞) in the continuous space. The first two derivatives ofgφ(bφ)are
∂gφ
∂bφ(bφ) =cs,φ+1
2(ch,φ+ch,J+1)− cls,φ·λφ
(bφ+ 1)2, and ∂2gφ
∂2bφ(bφ) = 2cls,φ·λφ (bφ+ 1)3. Note that the second derivative is positive for bφ≥1. So the optimal base stock levelb∗φ for stationφwith demand rateλφ is obtained from
∂gφ
∂bφ(b∗φ) = 0∧b∗φ>1 or b∗φ= 1 =⇒ b∗φ= max
(s cls,φ·λφ
cs,φ+ 12(ch,φ+ch,J+1) −1,1 )
. According to Corollary 2.5.3 this single local minimum b∗φ is also a global minimum.
To simplify notation we fix locationφ1and compare the performance metrics and costs for arrival rateλφandpφwith the situation of reduced demand of rate bλφ1 :=λφ/N and only a portionpbφ1 :=pφ/N of items from the replenishment workstation being redirected to location φ1. The quantities related to the location φ1 will be tagged by a “b” and an index “φ1”. We havebγφ1 =γφ= 1. All other cost valuesbcls,φ1,bcs,φ1,bch,φ1 remain the same as cls,φ,cs,φ,ch,φ, and ch,J+1 are already fixed. Using the previous results the new optimal base stock levelbb∗φ
1 for the location with reduced demand is bb∗φ1 = max
(s cls,φ·λφ/N
cs,φ+ 12(ch,φ+ch,J+1) −1, 1 )
. (2.6.1)
Comparing the optimal base stock levels b∗φ to bb∗φ
1, we see that b∗φ ≥ bb∗φ
1. To be more precise,
if bb∗φ1 >1, then bb∗φ1 + 1 b∗φ+ 1 = 1
√
N, and if bb∗φ1 = 1, then bb∗φ1 + 1 b∗φ+ 1 > 1
√
N. (2.6.2) This implies for sufficiently large b∗φ
bb∗φ1 ≈ b∗φ
√N > b∗φ
N, ifbb∗φ1 >1, and bb∗φ1 & b∗φ
√N > b∗φ
N, ifbb∗φ1 = 1. (2.6.3) Equation (2.6.3) says that whenever the demand is scaled down by1/N the optimal base stock level scales only with 1/√
N. This scaling is maintained for the standard costs which we consider (at least ifbb∗φ1 >1). This follows from substitutingbb∗φ1 into
bgφ1(bbφ1) :=
cs,φ+1
2(ch,φ+ch,J+1)
·bbφ1 +cls,φ·bλφ1· 1 (bbφ1 + 1). We directly obtain forbb∗φ
1 >1 bgφ1(bb∗φ1)≈
cs,φ+1
2(ch,φ+ch,J+1) b∗φ
√
N +cls,φ·λφ
N · 1
(b∗φ+1)
√ N
≈ 1
√
N ·gφ(b∗φ) (2.6.4)
and forbb∗φ
1 = 1 bgφ1(bb∗φ1)&
cs,φ+1
2(ch,φ+ch,J+1) b∗φ
N +cls,φ·λφ·b∗φ
N & 1
N ·gφ(b∗φ). (2.6.5)
Remark 2.6.3. ≈ resp.&in (2.6.3) means the following:
a(b∗φ) :=
bb∗φ
1 + 1 b∗φ+ 1 −bb∗φ
1
b∗φ =
bb∗φ1 + 1
·b∗φ−bb∗φ1· b∗φ+ 1
b∗φ+ 1
·b∗φ
= b∗φ−bb∗φ
1
b∗φ+ 1
·b∗φ
= 1−bb
∗ φ1
b∗φ
b∗φ+ 1.
Becauseb∗φ≥bb∗φ
1 it follows that bb
∗ φ1
b∗φ ≤1. This implies
bb∗φ1+ 1 b∗φ+ 1 −bb∗φ1
b∗φ
=bb∗φ1+ 1 b∗φ+ 1 −bb∗φ1
b∗φ = 1−bb
∗ φ1
b∗φ
b∗φ+ 1 < 1 b∗φ+ 1. Hence, for every ε > 0 we have for all b∗φ ≥ max 1,1ε−1
that 0 < a(b∗φ) < ε which yields
bb∗φ
1
b∗φ+a(b∗φ) =bb∗φ
1 + 1 b∗φ+ 1
(= √1
N ifbb∗φ
1 >1,
> √1
N ifbb∗φ
1 = 1 ⇔ bb∗φ1
=
√1
N −a(b∗φ)
·b∗φ ifbb∗φ
1 >1,
>
√1
N −a(b∗φ)
·b∗φ ifbb∗φ1 = 1.
≈resp.&in (2.6.4) resp. (2.6.5) means the following:
Letε >0. Then we have for all b∗φ≥max 1,1ε−1
that0< a(b∗φ)< ε and get forbb∗φ
1 >1 bgφ1(bb∗φ1) =
cs,φ+1
2(ch,φ+ch,J+1) 1
√
N −a b∗φ
·b∗φ+cls,φ·λφ
N · 1
(b∗φ+1)
√ N
=
cs,φ+1
2(ch,φ+ch,J+1) 1
√N −a b∗φ
·b∗φ +
1
√
N −a b∗φ
·cls,φ·λφ· 1
b∗φ+ 1+a b∗φ
·cls,φ·λφ· 1 b∗φ+ 1
= 1
√
N −a b∗φ
cs,φ+1
2(ch,φ+ch,J+1)
·b∗φ+cls,φ·λφ· 1 b∗φ+ 1
!
+a b∗φ
·cls,φ·λφ· 1 b∗φ+ 1
= 1
√
N −a b∗φ
·gφ(b∗φ) +a b∗φ
·cls,φ·λφ· 1 b∗φ+ 1
2.6. Structural properties of the integrated system
and forbb∗φ
1 = 1 gbφ1(bb∗φ1) ≥
cs,φ+ 1
2(ch,φ+ch,J+1) 1
√N −a b∗φ
·b∗φ+cls,φ·λφ
N · 1
bb∗φ
1 + 1
b∗φ≥bb∗φ
≥ 1
cs,φ+ 1
2(ch,φ+ch,J+1) 1
√
N −a b∗φ
·b∗φ+ 1
N ·cls,φ·λφ· 1 b∗φ+ 1
≥
cs,φ+ 1
2(ch,φ+ch,J+1) 1
N −a b∗φ
·b∗φ +
1
N −a b∗φ
·cls,φ·λφ· 1
b∗φ+ 1+a b∗φ
·cls,φ·λφ· 1 b∗φ+ 1
= 1
N −a b∗φ
·gφ(b∗φ) +a b∗φ
·cls,φ·λφ· 1 b∗φ+ 1.
2.6.2.1. Pooling of general homogeneous locations: Refined numerical evaluation on the basis of product form structure
For the more general homogeneous system, we resort to a numerical investigation to verify whether equation (2.6.2) is still satisfied. To do so, we start with a system consisting of one location, then split it in two equal parts. Utilizing the separability and the decomposition property (2.5.1), we can reduce the problem to an isolated single location.
We consider the following fictional system with J = {φ}, pφ = 1, λφ = 1, cs,φ = 1, ch,φ = 2, cls,φ = 400, ch,J+1 = 1 and γφ ∈ [0.1,10]. The holding costs ch,J+1 at the central supplier are lower than the holding costs ch,φ at location φ. The very high cost cls,φ = 400 results from expensive items, which justify the base stock policy, as argued in the introduction. We chose these numbers to obtain sufficiently large optimal base stock levels. When the location is split, such that the difference between continuous and discrete version of b∗φis negligible.
The results are plotted in Figure 2.6.1, where “full demand” refers to the original system (one location with λφ= 1) and “partial demand” refers to one of the two split locations (which are identical, each with demandλbφ1 =bλφ2 = 1/2).
From Figure 2.6.1(a) we see that the values of b∗φ and bb∗φ
1 are highest, when γφ = 1.
Figure 2.6.1(b) demonstrates monotone decreasing behaviour of the cost functions gφ, and bgφ1 in γφ. Figure 2.6.1(c) shows that bb
∗ φ1+1 b∗φ+1 ≥ √1
2. The ratio is close to its lowest value when γφ= 1. That means that ifγφ is only slightly different from 1, we soon gain more than factor √1
2 when two locations are pooled.
If γφ deviates more from 1 we observe that bb
∗ φ1+1
b∗φ+1 ≈ 0.9. The consequence is that pooling two demand streams of equal rate λ2φ yields a reduction of the needed inventory by a factor close to 1/2.
This observation and conclusion is supported further by Figure 2.6.1(a). It is shown that for high replenishment rate ν = γφp·λφ
φ (with λφ = 1 fixed), the optimal base stock level at demand rate λφ and λ2φ are almost the same. Finally, from Figure 2.6.1(d) we conclude that pooling two identical locations in the homogeneous case for largeγφreduced the optimal total costs by a factor1/2, too.
γ1 2
4 6 8 10 12
0 1 2 3 4 5 6 7 8 9 10
b1* (full demand) b1* (partial demand)
(a)b∗1 andbb∗1
γ1 0
100 200 300
0 1 2 3 4 5 6 7 8 9 10
g1(b1*) (full demand) g1(b1*) (partial demand)
(b) g1(b∗1)andbg1(bb∗1)
γ1 0.70
0.75 0.80 0.85 0.90
0 1 2 3 4 5 6 7 8 9 10
(b1*+1)/(b1*+1) 1 2
(c) bbb∗1∗+1 1+1
γ1 0.5
0.6 0.7 0.8 0.9
0 1 2 3 4 5 6 7 8 9 10
g1(b1*)/g1(b1*) 1 2
(d) ggb1(bb∗1)
1(b∗1)
Figure 2.6.1.: The optimal base stock levels of the original system b∗1 and of the splitted systemsbb∗1 with corresponding optimal costsg1(b∗1)andgb1(bb∗1)from Section 2.6.2.
Summarizing: From pooling two homogeneous locations we can expect roughlyat least a gain of inventory reduction by a factor of1/√2, which is attained (approximately) in the 1-homogeneous case. Hence, for a subset of the parameter space, pooling is advisable, i.e. the “type-F-anomaly” [YS09, Section 2] does not occur.
Additional comments:
(1) The explicit product form expressions for the stationary distribution allow a more refined evaluation. We report only some interesting observations, which refer to an evaluation in the continuous optimization domain.
(i) Although Figure 2.6.1(a) suggests that the optimal base stock levels b∗φ and bb∗φ1 are maximal at γφ = 1, we want to stress that this is in general not the
case.
(ii) Although Figure 2.6.1(c) suggests that the quotientbb
∗ φ1+1
b∗φ+1 is minimal atγφ= 1, this is in generalnot the case as well.
(iii) Figure 2.6.1(c) suggests even more that √1
2 is always a lower bound for bb
∗ φ1+1 b∗φ+1. We performed a detailed numerical evaluation with parametersλφ= 1,pφ= 1, cs,φ,ch,φ,ch,J+1andνfrom{0.1,0.2, . . . ,0.9} ∪ {1,2, . . . ,9} ∪ {10,20, . . . ,100}
andcls,φ from{0,10,20, . . . ,1000}, which resulted in62,080,256different
sys-2.6. Structural properties of the integrated system
tem settings. They showed that the quotient bb
∗ φ1+1
b∗φ+1 fell below the value 0.99√
2
for less than 6% of the instances, and below 0.95√
2 for less than 0.5% of the instances. The smallest quotient was approximately 0.86√
2.
(2) The strong peaks in Figure 2.6.1 (near γφ = 1) are still waiting for an intuitive explanation. We emphasize that the possibility to detect these peaks strongly relied on the fact that due to the product form stationary distribution we could separate parts of the system (queues at locations) from other parts (inventory and central supplier).
2.6.3. Transformation of the stationary distribution We started with evaluation of π(n,k) =
Q
j∈Jξj(nj)
·θ(k), which made it easy to define and understand the cost structure of the system. Introducing later on for θ the isomorphicθ−(see (2.5.4) on page 32) offers additional valuable insight into the structure of the inventory-supplier part of the integrated system.
Consider a situation where the service times for production of raw material at the workstation of the central supplier are extremely long (by chance). Then it is intuitive that the on-hand inventory levels at all locations are low (and base stock levels are high to prevent stockouts). Similarly, for short replenishment production times we see high stock levels at all locations. Therefore, it is a tempting conjecture that the inventory levels are positively correlated.
θ−tells us that this intuition goes wrong in the long-run and in equilibrium: Inventory levels are independent for any fixed time point. Clearly, the inventory processes are not independent over time.