1.

Inventory and Supply

 Manufacturing
 Production
 Supply
 Supply network
 Service
 Design
 Control
 Model

 Inventory is fundamental to production (WHY?)

 Can we ever operate without inventory (WHY/WHY NOT?)
 Inventory level is controlled through ordering policy. The ordering policy needs to be optimized
by minimizing ordering and inventory holding costs while fulfilling external or internal demands.
 What happens if we do not fulfill demands?
 Each inventory control model has some assumptions and it is important to validate these
assumptions when applying a model to a real-life system.

1. Supply network

Supplier Manufacturer Distributor Retailer

Nature Customers
(external)

Stock Point Material Flow Internal customer Inventory

Purchase order: delivery w/o transformation

Production order: delivery with transformation

2. Costs
 Order (setup cost): occurs per order and independent of the order quantity
Communication costs, equipment setup, transportation
 Purchase, production costs: depends on the ordered quantity
Material, labor, energy
 Inventory costs: more inventory → larger costs
Opportunity costs (lot-size the opp. to make money by investment)
Insurance costs, storage cost
 WIP costs: major inventory
Backorder cost: Customer will wait
Stockout cost: customer will not wait → lot-size profit in addition

3. Order policy: when and how much to order → to minimize the costs need to be determined

Page 1 of 29
 Trade off
# 𝑺𝒊𝒛𝒆
𝑭𝒓𝒆𝒒𝒖𝒆𝒏𝒕 𝑶𝒓𝒅𝒆𝒓 𝒘𝒊𝒕𝒉 𝒔𝒎𝒂𝒍𝒍 𝒔𝒊𝒛𝒆 ↑ ↓
𝑳𝒆𝒔𝒔 𝒇𝒓𝒆𝒒𝒖𝒆𝒏𝒕 𝒐𝒓𝒅𝒆𝒓 𝒘𝒊𝒕𝒉 𝒍𝒂𝒓𝒈𝒆 𝒔𝒊𝒛𝒆 ↓ ↑

4. EOQ: basic model → demand is static

EPL → single stock point
Dynamic lot-size → demand is dynamic
Dynamic in supply network
Dynamic in closed/open system → Statistical models → multiple stock-points

(EOQ model) the most basic

1.1.1 Economic Order Quantity Model F.W. Harris: (1) How many parts to make at once;
(2) How much stock to keep at hand, Factory, The Magazine of Management, 1913

Deterministic, single stock point model

1.1.2 Model description

Seven Basic Assumptions: Delivery Zero lead time

1. Production (purchase) is instantaneous.

No WIP → don’t have to consider WIP cost

2. There is no capacity constraint.

Inventory capacity: storage capacity
Order capacity → no limit, capacity = ∞
Production capacity

3. Demand is deterministic. There is no uncertainty in the quantity or timing of demand.

Demand Example: → 500 units / year

Time
Note: If demand is deterministic and dynamic → dynamic lot-size
Changes over time
q

Dynamic lot-sizing time

4. Demand is constant over time.

Rate is fixed

Page 2 of 29
 5. A production (purchase) run incurs a fixed setup cost. Regardless of the size of the lot or the status
of the factory, the setup cost is the same.
Order 200 → $ 10 Usually, variable setup costs.
Order 500 → $ 10 (Setup costs = f(order size))

6. Products can be analyzed individually.

Generally many products and assume they are independent
But in reality they have some interactions.
Example:
 Share storage
 Share transportation → becomes very complex → simplification
 Share same setup
Optimal decisions for all products are dependent on each other

7. Backorders are not allowed.

(stockouts)
All demand is fulfilled

Notations, model behavior, and analysis:

D = demand rate (in units per year)

is deterministic and constant

c = unit production (purchase) cost, not counting setup or inventory costs (in dollars per unit)
This represents the value of a product, material, labor, energy + time, space + environment
Ordering cost
(Cost to keep a unit for a year)
A = fixed setup cost to produce (purchase) a lot [$]

h = inventory holding cost (in dollars per unit per year); if the holding cost consists entirely of interest
on money tied up in inventory, then h=ic, where i is the annual interest rate,

Time depended. Opportunity, insurance and storage, the money tied up with inventory could
be used to make money by another investment

Q = lot size (in units); this is the decision variable,

Order size, Order quantity, a.k.a batch size, lot size…

Inventory behavior of EOQ model in an optimal ordering policy:

The behavior of the system can be represented by the behavior of inventory which can be used to model
the cost.
Slope is constant (demand is constant)
Order
Quantity

Problem in this behavior

1. Can reduce inventory cost → lower peaks, should hit zero
2.

Page 3 of 29
Irregular
order
quantity
Regular (average costs)

The costs should be in the inside of somewhere;

Irregular pattern cannot be optimal

Therefore, this is the optimal pattern:

q
Q = TD
T = Q/D
D = 1000/year Q
Q = 500
T = 1/2 1 time
q

T A A A A time
Setup = A
In this pattern we want to find order quality Q*

Total cost per year, which we denote by Y(Q), can be expressed as

ℎ𝑄 𝐴𝐷
𝑌(𝑄) = + + 𝑐𝐷
2 𝑄

≡ Q/2

T = Q/D
Costs per each interval = Setup + Production + Inventory holding+ WIP + Backorder
Costs / year Y (Q)
𝐐 𝑸 𝑸𝟐
= 𝐀 + 𝐂𝐐 + 𝐱 𝒉 = 𝑨 + 𝑪𝑸 + 𝒉
𝟐 𝑫 𝟐𝑫

Page 4 of 29
 𝑸𝟐
𝐘(𝐐) = (𝐀 + 𝐂𝐐 + 𝒉) 𝐱 ⋕ 𝐨𝐟 𝐨𝐫𝐝𝐞𝐫𝐬/𝐲𝐞𝐚𝐫
𝟐𝑫
𝑸𝟐 𝐃
𝐘(𝐐) = (𝐀 + 𝐂𝐐 + 𝒉) 𝐱
𝟐𝑫 𝐐
𝐀𝐃 𝐐
𝐘(𝐐) = + 𝐂𝐃 + 𝐡
𝐐 𝟐
Now, Find Q* that minimizes Y (Q)

Unique minimum point

AD/Q Y(Q)
The lot size that minimizes Y(Q) in the cost function is
Cost hQ/2
∗ 2𝐴𝐷
𝑄 = √
ℎ
cD
𝒅𝒀(𝑸) 𝑨𝑫 𝒉
= − 𝑸𝟐 + 𝟐 = 𝟎
𝒅𝑸 Q* Q
𝟐𝐀𝐃
𝐐= ±√ Q>0
𝐡

𝟐𝐀𝐃
𝐐∗ = √ Independent of c → we fulfill all demand, annual production cost is constant
𝐡
𝟐𝐀𝐃
𝐡√ 𝐀𝐃
𝐡
𝐘(𝐐∗ ) = + + 𝐜𝐃 = √𝟐𝐀𝐃𝐡 + 𝐜𝐃 Optimal cost
𝟐 𝟐𝐀𝐃
√
𝐡

This square root formula is the well-known economic order quantity (EOQ), also referred to as the
economic lot size. And, optimal order interval is

2𝐴
𝑇∗ = √
ℎ𝐷

𝐐∗ √𝟐𝐀𝐃 𝟐𝐀
𝐡
𝐓∗ = = =√
𝐃 𝐃 𝐡𝐃

𝟐𝐀𝐃
𝐐∗ = √
𝐡

𝟐𝐀
Optimal inventory behavior: 𝐓 ∗ = √𝐡𝐃

Page 5 of 29
 𝐘(𝐐∗ ) = √𝟐𝐀𝐃𝐡 + 𝐜𝐃
Optimal policy: Order Q* at every T* interval

Example. Consider a shop that produces metal racks. The demand for metal racks is fairly steady and
predictable at D=1,000 units per year. The unit cost of the racks is c=$250, but the metalworking shop
also charges a fixed cost of A=$500 per order, to cover the cost of shutting down the shop to set up.
The shop estimates its opportunity cost for money at 10 percent per year. It also estimates that the
floor space required to store a rack costs roughly $10 per year in annualized costs.

D = 1000 units /year

C = $ 250
A = $ 500
i = 10 % / year = 0.1
h = opportunity cost + storage cost
h = 0.1 x 250 + 10 = $ 35 / year
𝟐𝐀𝐃 𝟐 ∗ 𝟓𝟎𝟎 ∗ 𝟏𝟎𝟎
𝐐∗ = √ =√ = 𝟏𝟔𝟗 𝐮𝐧𝐢𝐭𝐬
𝐡 𝟑𝟓
𝐐∗ 𝟏𝟔𝟗
𝐓∗ = = = 𝟎. 𝟏𝟔𝟗 𝐲𝐞𝐚𝐫𝐬
𝐃 𝟏𝟎𝟎𝟎 ∗
𝒉𝑸 𝑨𝑫 𝟑𝟓 ∗ 𝟏𝟔𝟗 𝟓𝟎𝟎 ∗ 𝟏𝟎𝟎𝟎
𝒀∗ (𝑸) = + ∗ + 𝒄𝑫 = + + 𝟐𝟓𝟎 ∗ 𝟏𝟎𝟎𝟎 = $𝟐𝟓𝟓, 𝟗𝟏𝟔
𝟐 𝑸 𝟐 𝟏𝟔𝟗
Discussions:
 Production cost c does not affect the optimal policy
Y(Q) = + +cD (constant)
Y’(Q) = 0
Q* T* independent of c

 Consider a fixed lead-time L

A’ Ordered quantity is available at the same time after L.

Unit WIP cost/year = h’
Annual WIP cost = D x L x h’ (constant) (not affected by Q)
So, in this case
Q* is the same, only order point is different

1.1.3 Sensitivity

What do we mean by sensitivity?

𝑸∗ → 𝑸′ 𝒀(𝑸′ )
↓ ↓ ⇒ 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦
𝒀(𝑸∗ )
𝒀(𝑸∗ ) 𝒀(𝑸′ )

Page 6 of 29
If the rate is much less than the expected, we can say that the cost is insensitive to lot size.

The sum of holding and setup costs is fairly insensitive to lot size. To examine the sensitivity of the cost to
lot size, we formulate optimal total cost per year Y* (but omitting the c term, since this is not affected by
lot size).
(Y constant) make the analysis by analyzing production cost.

ℎ𝑄 ∗ 𝐴𝐷
𝑌 ∗ = 𝑌(𝑄 ∗ ) = + ∗ = √2𝐴𝐷ℎ
2 𝑄

𝒉𝑸 𝑨𝑫
𝒀(𝑸) = + + 𝒄𝑫
𝟐 𝑸
𝟐𝐀𝐃
𝐐∗ = √
𝐡
𝟐𝐀𝐃
𝒉√ 𝑨𝑫 √𝟐𝐀𝐃𝐡 √𝟐𝐀𝐃𝐡
𝐡
𝒀(𝑸∗ ) = + = + = √𝟐𝐀𝐃𝐡
𝟐 𝟐𝐀𝐃 𝟐 𝟐
√
𝐡
Optimal policy cost with Q* w/o production cost (it is not the total annual cost!)

Now, suppose that instead of using Q*, we use some other arbitrary lot size Q’, which might be larger or
smaller than Q*.

ℎ𝑄 ′ 𝐴𝐷
𝑌(𝑄′) = + ′
2 𝑄

Then,

𝑌(𝑄′) 1 𝑄 ′ 𝑄 ∗
= ( + )
𝑌(𝑄 ∗ ) 2 𝑄 ∗ 𝑄 ′

𝒉𝑸′ 𝑨𝑫
𝟐 + 𝑸′ 𝒉𝑸′ 𝑨𝑫 √𝒉𝑸′ √𝑨𝑫 𝟏 𝒉 𝟐𝑨𝑫 𝟏
= + = + = (√ 𝑸′ + √ )
√𝟐𝐀𝐃𝐡 𝟐√𝟐𝐀𝐃𝐡 𝑸′√𝟐𝐀𝐃𝐡 𝟐√𝟐𝐀𝐃 𝑸′√𝟐𝒉 𝟐 𝟐𝑨𝑫 𝒉 𝑸′
∗
𝟏 𝑸′ 𝑸
= ( + )
𝟐 𝑸∗ 𝑸′

Lot size sensitivity

To appreciate that, suppose Q’=2Q*, which implies that we use a lot size twice as large as optimal. Then,
the ratio of the resulting holding plus setup cost to the optimum is 1/2(2+1/2) = 1.25. That is, a 100 percent
error in lot size results in a 25% error in cost. Notice that if Q’=Q*/2, we also get an error of 25 percent in
the cost function.

Q’ = 2Q* (error in lot size)

Page 7 of 29
 → →

𝒀(𝑸′) 𝟏 𝟐𝑸∗ 𝑸∗ 𝟏 𝟏
∗
= ( ∗
+ ∗
) = (𝟐 + ) = 𝟏. 𝟐𝟓
𝒀(𝑸 ) 𝟐 𝑸 𝟐𝑸 𝟐 𝟐

Take double lot size →25% error (increase) in cost →cost is “insensitive” to lot size
Half in lots-size

𝟏 ∗
′
𝟏 ∗ 𝟏 𝟐𝑸 𝑸∗ 𝟏 𝟏
𝑸 = 𝑸 = ( ∗ + ) = ( + 𝟐) = 𝟏. 𝟐𝟓
𝟐 𝟐 𝑸 𝟏 ∗ 𝟐 𝟐
𝑸
𝟐

Half Q*→only 25% increase in cost

Order time sensitivity

We can get further sensitivity insights from the EOQ model with respect to order interval. Similar to the
previous case, the ratio of the cost resulting from an arbitrary order interval T’ and the optimal cost:

𝐴𝑛𝑛𝑢𝑎𝑙 𝑐𝑜𝑠𝑡 𝑢𝑛𝑑𝑒𝑟 𝑇′ 1 𝑇 ′ 𝑇 ∗

= ( + )
𝐴𝑛𝑛𝑢𝑎𝑙 𝑐𝑜𝑠𝑡 𝑢𝑛𝑑𝑒𝑟 𝑇 ∗ 2 𝑇 ∗ 𝑇 ′

T*→Y(T*) Q
T’→Y(T’)
𝒀(𝑻′ )
𝒀(𝑻∗)
≥𝟏 Q’
∗
𝑸
𝑻∗ =
𝑫
𝑸
𝑻 = 𝑫 → 𝑸 = 𝑻𝑫 → 𝑻 ↔ 𝑸
T’ T
𝒀(𝑻∗ ) = √𝟐𝑨𝑫𝒉

𝒉𝑸′ 𝑨𝑫
𝒀(𝑻′ ) = 𝟐
+ 𝑸′
where Q’ = T’ D
𝒉𝑻′𝑫 𝑨𝑫 𝒉𝑻′𝑫 𝑨
𝒀(𝑻′ ) = + = +
𝟐 𝑻′𝑫 𝟐 𝑻′
𝒉𝑻′𝑫 𝑨
𝒀(𝑻′ ) 𝟐 + 𝑻′ 𝑻′√𝒉𝑫 √𝑨 𝟏
√
𝒉𝑫 ′
√
𝟐𝑨 𝟏 𝟏 𝑻′ 𝑻∗
= = + = ( 𝑻 + )= ( ∗+ )
𝒀(𝑻∗ ) √𝟐𝑨𝑫𝒉 𝟐√𝟐𝑨 𝑻′√𝟐𝒉𝑫 𝟐 𝟐𝑨 𝒉𝑫 𝑻∗ 𝟐 𝑻 𝑻′
Ex: T’ = 2 T*

Page 8 of 29
 T’= ½ T*
Only 25% increase in cost
𝟏 𝟏
(𝟐 + ) = 𝟏. 𝟐𝟓
𝟐 𝟐

Now, suppose that the optimal order interval T* is represented in weeks, and lies between 2m and 2m+1 for
some integer m. We choose 2m if T* lies in the interval [2m, 2m√2] and 2m+1 otherwise. Then, the error is
guaranteed to be no more than 6%.

𝐴𝑛𝑛𝑢𝑎𝑙 𝑐𝑜𝑠𝑡 𝑢𝑛𝑑𝑒𝑟 𝑇′

≤ 1.06
𝐴𝑛𝑛𝑢𝑎𝑙 𝑐𝑜𝑠𝑡 𝑢𝑛𝑑𝑒𝑟 𝑇 ∗

2m ≤ T* ≤ 2m+1 m is an integer

T* = 5 weeks
22 ≤ T* ≤ 23 →m=2
If 2m ≤ T* ≤ √𝟐 2m → T’ = 2m choose T’ as order interval
If √𝟐2m ≤ T* ≤ 2m+1 → T’ = 2m+1 → 4….5.66………8 → T’ = 4
Then the question is how much the cost change is, if we apply this ordering rule.
Case 1
2m ≤ T* ≤ √𝟐 2m → T’ = 2m
𝒀(𝑻′ ) 𝟏 𝟐𝒎 𝑻∗ Cost
𝒀(𝑻∗)
= 𝟐 ( 𝑻∗ + 𝟐𝒎 )

(T*)2 = (2m)2
T* = 2m
𝟏 𝟐𝒎 𝟐𝒎 T*
If T* = 2m → ( 𝒎 + 𝒎) = 𝟏 T
𝟐 𝟐 𝟐
𝒎
𝟏 𝟐 √𝟐𝟐𝒎 𝟏 𝟏
If T* = √𝟐2m → 𝟐 (√𝟐𝟐𝒎 + 𝟐𝒎 ) = (√𝟐 +
𝟐 √𝟐
) = 𝟏. 𝟎𝟔
2m 2m+1
𝒀(𝑻′ )
𝒀(𝑻∗)
≤ 𝟏. 𝟎𝟔 → error is less than 6%

Case 2
√𝟐2m ≤ T* ≤ 2m+1 → T’ = 2m+1
𝒀(𝑻′ ) 𝟏 𝟐𝒎+𝟏 𝑻∗
= ( + )
𝒀(𝑻∗ ) 𝟐 𝑻∗ 𝟐𝒎+𝟏
T* = 2m+1
𝟏
If T* = 2m+1 → 𝟐 (𝟏 + 𝟏) = 𝟏
𝟏 𝟏
If T* = √𝟐2m → 𝟐 (√𝟐 + ) = 𝟏. 𝟎𝟔
√𝟐
′
𝒀(𝑻 )
< 1. 𝟎𝟔
𝒀(𝑻∗ )

All together if we follow this rule

Page 9 of 29
𝒀(𝑻′ )
≤ 𝟏. 𝟎𝟔
𝒀(𝑻∗ )
(6% maximal increase in cost (Under this policy, an order is released every 2m
weeks with some integer m)

This approach is called optimal power-of-2 policy. This result is useful in multiproduct settings, where it is
desirable to order such that different products are frequently replenished at the same time (e.g. to facilitate
sharing of delivery trucks). That is, make the order interval 1 week, 2 weeks, 4 weeks, 8 weeks, and so
forth. The result is that items ordered at 2n-week intervals will be placed at the same time as orders for items
with 2k intervals for all k smaller than n. This will facilitate sharing of trucks, consolidation of ordering
efforts, simplification of shipping schedules. If the increases in holding and order costs of all the items (no
more than 6%) are offset by the reduction of such costs, the power-of-2 order schedule is worthwhile.

1. Consider that we have multiple products and we follow this policy, for example considerer 3
products.
Product 1 → T’ = 2
Product 2 → T’ = 4
Product 3 → T’ = 8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

A product ordered every 2n weeks is ordered at the same time as those with 2k intervals when
k≤n
23 → 22 2’
22 → 2’

2. Benefits of group orderly: sharing delivery trucks

sharing common setups
Simplifications of shipping schedule
Not easy to optimizes consider all the products simultaneous, this is practically impossible

So, we are getting many benefits at the expense of 6% cost income in the worst case.

Example. In the previous example, Q* = 169 and T* = Q*/D = 0.169 year.

T* = 0.169 years = 52 x 0.169 weeks = 8.78 weeks

23 ≤ T* ≤ 24
23 (8) (8.78) √𝟐 23 (11.31) 23 (16)

T’ = 23 order internal
T’ = 8 weeks = 8/52 years = 0.15 x years

𝒀(𝑻′ ) 𝟏 𝑻′ 𝑻∗ 𝟏 𝟖. 𝟕𝟖 𝟖
= ( + ) = ( + ) = 𝟏. 𝟎𝟎𝟒 ≪ 𝟏. 𝟎𝟔
𝒀(𝑻∗ ) 𝟐 𝑻∗ 𝑻′ 𝟐 𝟖 𝟖. 𝟕𝟖

Page 10 of 29
(0.4% error in cost)

Discussions:

1. Optimal power of this policy (n ≥2)

nm ≤ T* ≤ nm+1
If nm ≤ T* ≤ √𝒏nm → T’ = nm
If √𝒏nm ≤ T* ≤ nm+1 → T’ = nm+1

𝒀(𝑻′) 𝟏 𝟏
∗
≤ (√𝒏 + )
𝒀(𝒀 ) 𝟐 √𝒏

2. A (is actual), h, D, c → estimation error

A’ = estimation

𝟐𝑨′𝑫
𝑸′ = √
𝒉
That this is optimal
𝒀∗ = √𝟐𝑨𝑫𝒉
𝟐𝑨′𝑫
𝒉√ 𝑨𝑫
𝒉
𝒀′ = + →Actual reactor cost
𝟐 𝟐𝑨′𝑫
√
𝒉

𝟐𝑨′𝑫
𝒉√
𝒉 + 𝑨𝑫
𝟐
𝒀′ √𝟐𝑨′𝑫 𝟏 𝑨′ 𝑨
𝒉 √ +√ )
= = (
𝒀∗ √𝟐𝑨𝑫𝒉 𝟐 𝑨 𝑨′
𝟏 𝟏
𝑨′ = 𝟐𝑨 = (√𝟐 + ) = 𝟏. 𝟎𝟔
𝟐 √𝟐
𝟏 𝟏 𝟏
𝑨′ = 𝑨 = (√𝟐 + ) = 𝟏. 𝟎𝟔
𝟐 𝟐 √𝟐

6% error in cost is insensitive

We can do the same thing for other parameters.

1.1.3 EOQ extensions

One extension of EOQ model is to the case in which replenishment is not instantaneous; instead, there is a
finite, but constant and deterministic, production rate. This model is sometimes called the economic
production lot (EPL) model. If we let P represent the production rate (and assume that P > D so that the
system has capacity to keep up with demand), then the EPL model results in the following lot size to
minimize the sum of setup and holding costs:
(P < D cannot fulfill demand) (P
=D continually produces)

Page 11 of 29
 2𝐴𝐷
𝑄∗ = √
ℎ(1 − 𝐷/𝑃)

Note that if P=∞, then the product is instantaneity; this formula reduces to the regular EOQ. Otherwise, it
results in a larger lot size to cover for the fact that replenishment items take time to produce. Other variants
of the basic EOQ include backorders (i.e., orders that are not filled immediately, but have to wait until stock
is available), major and minor setups, and quantity discounts among other.
An optimal pattern:
consumption

product inventory level

(P-D) x t1
Q is demanded with rate D
P-D D T x D = Q fact 1
𝑸
𝑻=
𝑫
T

t1 t2
Here Q is demanded with rate P

Q = t1 x P fact 2
𝑸
𝒕𝟏 =
𝑷
𝑸 𝑸
𝒕𝟐 = 𝑻 − 𝒕 𝟏 = −
𝑫 𝑷
Inventory level
𝑸
𝒕𝟏 𝒙(𝑷 − 𝑫) = (𝑷 − 𝑫)
𝑷
𝟏𝑸 𝑸 𝑫 𝑨𝑫 𝑸(𝑷 − 𝑫)
𝒀(𝑸) = [𝑨 + 𝒄𝑸 + (𝑷 − 𝑫) 𝒉] = + 𝒄𝑫 + 𝒉
𝟐𝑷 𝑫 𝑸 𝑸 𝟐𝑷
𝝏𝒀(𝑸) 𝑨𝑫 (𝑷 − 𝑫)
=− 𝟐 + 𝒉=𝟎
𝝏𝑸 𝑸 𝟐𝑷
𝑨𝑫 𝑷 − 𝑫
= 𝒉 → (𝑷 − 𝑫)𝒉𝑸𝟐 = 𝟐𝑷𝑨𝑫
𝑸𝟐 𝟐𝑷

Page 12 of 29
 𝟐𝑷𝑨𝑫 𝟐𝑨𝑫
𝑸∗ = √ =√
(𝑷 − 𝑫)𝒉 𝒉(𝟏 − 𝑫/𝑷)

Economic production lot

Other models
1 EOQ + Backorder
S
Order quantity
Two decision variables
Q Q.q→S=Q-q

q back order quantity

Inventory represents the amount of backorder

2 EPL + Backorder

Order

3 EOQ + Stockout

Negative = amount of stockouts

Q
Variations Models can be formed and analyzed
q
Lost sales

4 EPL + Stockout

Page 13 of 29
 Seasonal demands. E.g., umbrella, (Demand is constant)
1.2 Dynamic Lot Sizing refrigerator, fan
When demand varies over time, a continuous time model, like the EOQ model, is awkward to specify. So,
instead, we will clump demand into discrete time periods, which could correspond to days, weeks, or
months, depending on the system. A daily production schedule might make sense for a high-volume system
with rapidly changing demand, while a monthly changing schedule may be adequate for a low-volume
system with demand that changes more slowly.
1 Forecast demands for future periods (days, weeks and month) using some forecasting
technique

Demand over planning horizon

planning horizon

2 Solve an optimization problem in terms of optimal order schedule.

3 Release only the optimal order for the current time

4 In the next periods, same process is repeated.

This is called Rolling horizon control (RHC) and is popular in practice or in model predictive
control (MPC) (general approach in many areas other than Inventory control).

In RHC two parameters:

Size of period Planning Horizon
(day or week) (year)
Forecasting cost (↑) ↑ ↓
Forecast accuracy (↑) ↓ ↑
Production demand volume (↑) ↓ ↑

1.2.1 Model description

In this context we discuss the optimal problem. Determine optimal order schedule over a planning
horizon.

Assumptions: no lead time, no backorder

Notations:

t = a time period (we will assume weeks, but any interval could be used); the range of time periods is
t = 1, …, T, where T represents the planning horizon
Dt = demand in week t (in units)
ct = unit production cost (in dollars per unit), not counting setup or inventory cost in week t
At = setup cost to produce (purchase) a lot in week t (in dollars)
ht = inventory holding cost to carry a unit of inventory from week t to week t+1 (in dollars per unit per
week);
(Per period, in week t)
It = inventory (in units) leftover at the end of week t

Page 14 of 29
 Qt = lot size (in units) in week t; there are T such decision variables, one for each week
(but we are using I1 for simplification of formulas)

I0 I1 I2 It IT

Q1 Q2 Q3 Qt QT
! 2 3 t … T

D1 D2 D3 Dt DT
C1 C2 C3 Ct CT
A1 A2 A3 At AT
h1 h2 h3 ht hT

I0 → quantitative inventory level

D, A, C, h → 1st parameters
I, Q, T →2nd variables
QT –→at the beginning of T = at the end of T-1
Dt → demand rate in each period is assumed constant

Inventory balance equation:

I t 1  Qt  Dt  I t t  1,2,.., T

Qt period t

It-1 Dt It
Qt (no lead time)

I balance equations
t=1 I0 + Q1 - D1 = I1
t=2 I1 + Q2 – D2 = I2
t=t It-1 + Qt – Dt = It
…
t=T IT-1 + QT – DT = IT
D1, D2, parameters, not variables

We want to minimize the total cost over the planning horizon

Cost: setup cost: At if Qt > 0

Production cost: CtQt
Inventory cost: htIt → conventionally → we talk about this later.
Some simple rules

Lot-for-lot rule: the simplest lot-sizing procedure one might think of is to produce exactly what is
required in each week.

Page 15 of 29
 t 1 2 3 4 5 6 7 8 9 10 Total
Dt 20 50 10 50 50 10 20 40 20 30 300
Qt 20 50 10 50 50 10 20 40 20 30 300
It 0 0 0 0 0 0 0 0 0 0 0
Setup cost 100 100 100 100 100 100 100 100 100 100 1000
Holding cost 0 0 0 0 0 0 0 0 0 0 0
Total cost 100 100 100 100 100 100 100 100 100 100 1000
At=$100 for all t, ht =$1 for all t, ct=constant for all t, I0 = 0

No inventory cost
Setup in every period → too much setup cost

Fixed order quantity rule: Another plausible policy is to produce a fixed amount each time we perform a
setup.

t 1 2 3 4 5 6 7 8 9 10 Total
Dt 20 50 10 50 50 10 20 40 20 30 300
Qt 100 0 0 100 0 0 100 0 0 0 300
It 80 30 20 70 20 10 90 50 30 0 0
Setup cost 100 0 0 100 0 0 100 0 0 0 300
Holding cost 80 30 20 70 20 10 90 50 30 0 400
Total cost 180 30 20 170 20 10 190 50 30 0 700
At=$100 for all t, ht =$1 for all t, ct=constant for all t, I0 = 0

Total demand = 300

Suppose we order 3 times each with 100 units
Alternative
We can use 50 units → 6 times, or 60 units → 5 times, etc.
10 times is the maximum allowed, then what is the best fixed order quantity?
t=1
Dt = 60
Qt = 50
Some of these may not be feasible

IP formulation:

Min ∑𝑻𝒕=𝟏[𝒔𝒆𝒕𝒖𝒑 + 𝒑𝒓𝒐𝒅𝒖𝒄𝒕𝒊𝒐𝒏 + 𝒊𝒏𝒗𝒆𝒏𝒕𝒐𝒓𝒚]

𝑻

∑[𝑨𝒕 𝒚𝒕 + 𝒄𝒕 𝑸𝒕 + 𝒉𝒕 𝑰𝒕 ]
𝒕=𝟏
yt = 0 if Qt = 0
yt = 1 if Qt > 0

s.t.
1. Qt ≤ Myt t = 1,2,…,T

M is a large number
If Qt = 0 yt → 0 or 1 →minimize problem, so yt will be 0

Page 16 of 29
 If Qt > 0 yt = 1

2. It-1 + Qt – Dt = It t = 1, …, T

3. Qt ≥ 0, It ≥ 0 (no back orders) t = 1, …, T

4. yt = 0 or 1 t = 1, …, T
Just in case
M ≥ Σ Dt Think about this

Why?
Then you do not limit your solution space
You are considering all possible/ feasible solutions
Consider the case when you order all at t = 1
We used ΣhtIt for inventory cost. Let’s examine this.
Assume that demand is constant in each period.
Inventory behavior:
I0

Q1
I1

1 2 T

𝑰𝟎 + 𝑸𝟏 + 𝑰𝟏 𝑰𝟏 + 𝑸𝟐 + 𝑰𝟐 𝑰𝑻−𝟏 + 𝑸𝑻 + 𝑰𝑻
= 𝒉𝟏 + 𝒉𝟐 + ⋯ + 𝒉𝑻
𝟐 𝟐 𝟐
𝒉𝟏 𝑰𝟎 (𝒉𝟏 + 𝒉𝟐 )𝑰𝟏 (𝒉𝟐 + 𝒉𝟑 )𝑰𝟐 (𝒉𝑻−𝟏 + 𝒉𝑻 )𝑰𝑻−𝟏 𝒉𝑻 𝑰𝑻 𝒉𝟏 𝑸𝟏 + 𝒉𝟐 𝑸𝟐 + ⋯ + 𝒉𝑻 𝑸𝑻
= + + + ⋯+ + +
𝟐 𝟐 𝟐 𝟐 𝟐 𝟐
Suppose, ht = h for all t
𝒉𝑰𝟎 𝒉𝑰𝑻 𝒉 ∑ 𝑸𝒕
= + 𝒉𝑰𝟏 + 𝒉𝑰𝟐 + ⋯ + 𝒉𝑰𝑻−𝟏 + +
𝟐 𝟐 𝟐

Constant actual point of minimal It = 0 ∑ 𝑸𝒕 = ∑ 𝑫𝒕 ≡ 𝐌𝐢𝐧(∑ 𝐡𝐈𝐭 + ⋯)

When we minimize can give you same solution

So when ht = h the IP solution is exact
But when ht ≠ h the IP solution is approximate

Thomson M. Whitin (1923 –2013); Harvey M. Wagner (1931 – July 2017)

American management scientists; coauthored:
Dynamic Problems in the Theory of the Firm, 1957. 44 years after EOQ
1.2.2 Wagner-Whitin Procedure Management Science, 1958
~60 years ago
 The exact solution by IP is computationally impractical
 Not a heuristic but exact algorithm Whitin ~34; Wagner ~26

Page 17 of 29
A key observation for solving the dynamic lot-sizing problem is that if we produce items in week t (and
incur a set up cost) for use to satisfy demand in week t+1, then it cannot possibly be economical to produce
in week t+1 (and incur another setup cost). Either it is cheaper to produce all of week t+1’s demand in week
t, or all of it in t+1; it is never cheaper to produce some in each. (Notice that we violated this property in
the fixed order quantity rule). In more general terms, we can state this result as follows:
It-1 = 0
Dt = 50 Dt+1 = 100 Qt = 70 Qt+1 = 80
It = 20 It+1 = 0
t t+1

It is always better to produce

Qt = 50 Qt+1 = 100
It = 0 It+1 = 0
Same setup but less inventories cost
Partial production for next period cannot be optimal if you have two set-ups

Also consider this schedule

Qt = 150 Qt+1 = 0
It = 100 It+1 = 0
Just one setup but large inventory cost
No partial production (produce all for current + next periods)
Either one can be optimal, depending on the tradeoff between setup and inventory cost.

Wagner-Whitin Property: Under an optimal lot-sizing policy either the inventory carried to week t+1 from
a previous week will be zero or the production quantity in week t+1 will be zero.

General property of optimal solution structure.

t t+1

Either It = 0 (no inventory can be carried to next period)

Or Qt+1 = 0 (no production in t+1)
At least one of them should be zero; cannot be both positive
Note: Lot – for – lot and fixed order quantity can work, but may not be optimal

The Wagner-Whitin property greatly facilitates computation of optimal production quantities. It implies
that either Qt=0 or Qt will be exactly enough to satisfy demand in the current week plus some integer number
of future weeks. We could compute the minimum-cost production schedule by enumerating all possible
combinations of weeks in which production occurs. However, since we can either produce or not produce
in each week, the number of such combinations is 2T-1, which can be quite large if many weeks are
considered. To be more efficient, Wagner and Whitin (1958) suggested an algorithm that is well suited to
computer implementation.

If WW property holds, optimal order schedule should be a group ordering schedule

→ satisfy WW property

Page 18 of 29
 Qt > 0 It > 0
It-1 = 0 Qt+1 = 0

Also only this group ordering satisfies WW property

Size of solution space (that satisfies WW property)
0 0 0 … 0
1 1 1 … 1 2T-1 order at t=1
Each combination results in a group ordering schedule and satisfies WW property
Overall Procedure:

The solution space is large T>0

(DP + WW property) efficient algorithm
Assume I0 = 0 but can be modified easily for I0 > 0

Zt*: optimal cost of first t period problem

1 2 t T

Z1 Z2 Zt* ZT*
1 period 2 periods t periods So, we are solving T Zt*
Problem problem problem problems

At time t: following WW; property

Alt 1 → produce Dt in t → It-1 = 0 Optimal cost

Zt-1* At + ctDt + 0 + Zt-1*

Independent prob

Alt 2 → produce Dt in t-1 → It-2 = 0 → Dt-1 + Dt

Zt-2* At-1 + ct(Dt-1 + Dt) + ht-1Dt + Zt-2*

Alt 3 → produce Dt in t-2 → It-3 = 0 → Dt-2 + Dt-1 + Dt

Zt-3* At-2 + ct(Dt-2 + Dt-1 + Dt) + ht-2(Dt-1 + Dt) + ht-1Dt + Zt-3*

Case over inventory > 0, so no product at t-1 and t

Alt t → produce Dt in 1

Once we find Z1* → Z2* → … → ZT*

Then from this we can find the optimal schedule.

Step 1: We begin the algorithm by looking at the 1-week problem. The optimal policy for this problem is
trivial; we produce 20 units to satisfy demand in week 1, and we are done. Since there is no inventory
carried from one week to another, and we are neglecting production cost, the minimum cost in the 1-
week problem, which we denote by Z1*, is

Page 19 of 29
 Z1* = A1 = 100

The last week of production in the 1-week problem, which we denote by j1*, is

j1* = 1

Neglecting production cost because ct = constant = c

w/o production cost w/production cost → ctΣDt → same in optimal schedule, so, we only have the
minimum setup + inventory

At = $100 ht = $1 1
Only one choice → produce 20 D1 = 20
Z1* = A1 = 100
jt* : last week of production in t – week problem.
jt* = 1

Step 2: In the next step of the algorithm we increase the time horizon and consider the 2-week problem.
Now we have two options for the production in week 2; we can cover demand in week 2 with
production either week 1 or in week 2. If we produce if in week 1, we will incur a holding cost
associated with carrying inventory from week 1 to week 2. If we produce it in week 2, we will incur
an extra setup cost in week 2. Notice also that if we produce in week 2, then the cost of satisfying
previous demand is given by Z1*. Since we are trying to minimize cost, the optimal policy is to choose
the week with the lower total cost, that is,

Z2* = min {A1 + h1D2, Z1* + A2}

= min {100+1(50) = 150, 100+100 = 200)
=150

The optimal decision is to produce for both weeks 1 and 2 in week 1. Therefore, the last week in
which production takes place in an optimal 2-week policy is

j2* = 1 1 2
20 50

Z2* = min → optimal cost when producing D2 in week 2 (case 1)

→ optimal cost when producing D2 in week 1 (case 2)

Case 1

I1 = 0
Z1* + A2 + 0 = 100 + 100 = 200
Independent problem 1 week problem + optimal cost of 1-week Z1*

Case 2

50
A1 + I1h1 = 100 + 50 = 150

Page 20 of 29
Step 3: Now, we proceed to the 3-week problem. Ordinarily four possible production schedules would need
to be considered: produce in week 1 only, produce in weeks 1 and 2, produce in weeks 1 and 3, or
produce in weeks 1, 2, and 3. However, we need to consider only three of these: one only, one and
two, and one and three. This is because we need to consider only when we are going to produce the
demand for week 3. We have already solved the 2- and 1-week problems. If we decide to produce in
week 3, then we know from our solution to the 2-week problem that it will be optimal to produce for
weeks 1 and 2 in week 1. The savings in computation from this observation grow sharply as the
number of weeks grows.

Z3* = min {A1 + h1D2 + (h1+h2)D3, Z1* + A2 + h2D3, Z2* + A3}

= min {100 + 1(50) + (1+1)(10) = 170, 100 + 100 + 1(10) = 210, 150 + 100 = 250}
= 170

Again, it is optimal to produce everything in week 1, so

j3* = 1 1 2 3
20 50 10
Z3* = min → optimal cost when producing D3 in week 3 (case 1)
→ optimal cost when producing D3 in week 2 (case 2)
→ optimal cost when producing D3 in week 1 (case 3)

Case 1
Z2*
I2 = 0
Z2* + A3 + 0 = 150 + 100 = 250

Case 2

Z1*
I2 = 10 because I1 = 0

Z1* + A2 + I2h2 = 100 + 100 + 10 * 1 = 210

Case 3

Z1*
I1 = 60 I2 = 10

A1 + I1h1 + I2h2 = 100 + 60 + 10 = 170

Step 4: The situation changes when we move to the next step, the 4-week problem. Now, there are four
options for the timing of production for week 4, namely weeks 1 to 4:

Z4* = min {A1+h1D2+(h1+h2)D3+(h1+h2+h3)D4, Z1*+A2+h2D3+(h2+h3)D4, Z2*+A3+h3D4, Z3*+A4}

= min{100+1(50)+(1+1)(10)+(1+1+1)(50)=320, 100+100+1(10)+(1+1)(50)=310,
150+100+1(50) = 300, 170+100=270}
=270

Page 21 of 29
 This time, it turns out to be optimal not to produce in week 1, but rather to meet week 4’s demand
with production in week 4. Hence,

j4* = 4 1 2 3 4
20 50 10 50
Z4* = min
Z3* + A4 = 170 + 100 = 270
Z2* + A3 + I3h3 = 150 + 100 + 50 = 300
Z1* + A2 + I2h2 + I3h3 = 100 + 100 + 60 + 50 = 310
A1 + I1h1 + I2h2 + I3h3 = 100 + 110 + 60 + 50 =320

If our planning horizon were only 4 weeks, we would be done at this point. We would translate our results
to a lot-sizing policy by reading the jt* values backward in time. The fact that j4* = 4 means that we would
produce D4 = 50 units in week 4. This would leave us with a 3-week problem. Since, j3*=1, it would be
optimal to produce D1+D2+D3 = 80 units in week 1.

Z1* → Z2* → Z3* → Z4* → optimal

j1* → j2* → j3* → j4* → optimal schedule

suppose our problem is just 4-week problem (not 10 week)

Then, let’s see what the optimal schedule is
1 2 3 4
20 50 10 50
j1 = 1
* j2 = 1
* j3 = 1
* j4* = 4
For the 4-week problem it is optimal to produce in week 4
For the 3-week problem it is optimal to produce in week 1

80 independent 50 I3 = 0
Optimal schedule →Z4* = 270

Step 5: But, our planning horizon is not 4 weeks; it is 10 weeks. Hence, we must continue the algorithm.
However, before doing this, we will make an observation that will further reduce the computations
we must make.
We can reduce the computations by a property

Planning horizon property: If jt* = l, then, the last week in which production occurs in an optimal t+1 week
policy must be in the set l, l+1, …, t+1.

t t+1
jt* = l produce Dt+1
we only have to consider l~t+1
t+1
t t+1 alternatives → still we can find optimal schedule
t-1 optimal exist along these
l
…* don’t have to consider
jt+1* cannot be here, keys to find optimal schedule

Page 22 of 29
 Using this property, the calculation required to compute minimum cost for the 5-week problems is

Z5* = min {Z3*+A4+h4D5, Z4*+A5}

= min {170+100+1(50) = 320, 270 + 100 = 370}
= 320
Hence,

J5*=4
1 2 3 4 5
20 50 10 50 50
Z5* = min produce in 5 Z4* + A5 = 210 + 100 = 310
Produce in 4 Z3* + A4 + I4h4 = 170 + 100 + 50 = 320
X in 3
X in 2
X in 1
j5*=4
We solve the remaining 5 weeks, using the same approach. The results are in the table below.

Planning horizon t
1 2 3 4 5 6 7 8 9 10
1 100 150 170 320
2 200 210 310
3 250 300
4 270 320 340 400 560
5 370 380 420 540
6 420 440 520
7 440 480 520 610
8 500 520 580
9 580 610
10 620
Zt* 100 150 170 270 320 340 400 480 520 580
jt* 1 1 1 4 4 4 4 7 7 or 8 8
Either one is
fine, so, it is better to
choose 8 for less
computation
We are solving 10-week problem
It is optimal to produce week 8
J10* = 8

7 week problem
Week 1 = 80
Week 4 = 130
Week 8 = 90

Another example: same demands

JE* 1 2 2 2 4 6 6 8 8 9
→ order schedule

Page 23 of 29
Discussion:

1. Safety Stocks
We assumed deterministic demand, there can be uncertainly

Dt 99% target
Dt + S

We want to have this amount available in the beginning of each period.

It-1 + Qt
t

It

It-1 + Qt ≥ Dt + S
≡ It-1 + Qt - Dt ≥ S
≡ It ≥ S
So, when we consider safety stock.
same formulation but
It ≥ 0 replace this with It ≥ S

2. Lead time L = 3
1 2 3 4 5 T

I0 I1 I2 I3 I4 I5 …. IT

Q-1 Q0 Q1 QT-1

Q-1 , Q0, Q1 → constants, not decision variables

I0, I1, I2, I3 → constants
Min Σ[Atyt + ctQt + htIt]
s.t. It-1 + Qt-1 – Dt = It t
Qt ≤ Myt
Qt ≥ 0 t-L It-1 It
It ≥ 0
yt ϵ {0,1} Qt-1

Qt-L = constant
if t-L ≤ 0 ordered before but not received yet.
I0, I1, I2, …., Il all constants

Page 24 of 29
 3. Capacity constraints
Order capacity Qt ≤ a due to production limits, transportation limits
Inventory capacity It ≤ b storage limits

Joint capacity constraints

W1Qt + W2It ≤ C Human-power, electricity

Human-power for production Human-power for inventory management

1.3 Supply Networks [chains or networks?] -- Dynamic Lot-sizing

 Lot-sizing in a supply network, which is planning material flow through procurement, production,
and distribution to fulfill external demands, is a major driver of the costs and customer service.
 Because of its practical importance, the lot-sizing is among the most widely researched areas in
operational supply network management.
 When external demand pattern is dynamic, rolling horizon updating is a standard operational means
of incorporating new information available over time. The dynamic lot-sizing problem over a
planning horizon makes a plan that minimizes order and inventory costs, while satisfying forecasted
external demands.
 Since the stochastic nature of a supply network can be incorporated through the use of safety buffers
and rolling horizon updating, deterministic dynamic lot-sizing problem is a general framework for
the supply network planning.
 Though some problem models can be solved in polynomial time, in general, the dynamic lot-sizing
problem is hard.

Page 25 of 29
1.3.1 Background
Handled by ERP information systems, combining internal models of dynamic lot sizing.
Usually, solved independently by each supply network party.

Page 26 of 29
1.4 Closed-loop Supply Networks -- Dynamic Lot-sizing

Now, consider a closed-loop supply network as shown in the figure below. This system includes
remanufacturing, differently from the forward supply network we have discussed so far.

Forward
Nature Manufacturing

Customers

Remanufacturing
Serviceable
Inventory Used
products
Returns
Inventory
Backward
A simple model: Forward  Closed loop

Less production cost

Improved company image for social care, sustainability
Tax reduction
Green effect less consumption of natural resources
Less disposal to environment
But require information for collection and remanufacturing
(Needs cost – benefit analysis)

1.4.1 Model description

Assumptions: no backlogging, no lead time

Page 27 of 29
Page 28 of 29
Optional Reference: Factory Physics, Hopp and Spearman, 2011) Chapters 2 & 3
Further Reading (OPTIONAL) – Sample of Recent and Current Research Issues in this topic
area

Economic Order Quantity (EOQ) & Economic Production Lot (EPL)

[1] Oh, Y. H., & Hwang, H. (2006). Deterministic inventory model for recycling system. Journal of
Intelligent Manufacturing, 17(4), 423–428.
[2] Bouchery, Y., Ghaffari, A., Jemai, Z., & Dallery, Y. (2012). Including sustainability criteria into
inventory models. European Journal of Operational Research, 222(2), 229–240.
[3] Hua, G., Cheng, T. C. E., & Wang, S. (2011). Managing carbon footprints in inventory management.
International Journal of Production Economics, 132(2), 178–185.
[4] Chung, K.-J., & Cárdenas-Barrón, L. E. (2012). The complete solution procedure for the EOQ and
EPQ inventory models with linear and fixed backorder costs. Mathematical and Computer
Modelling, 55(11-12), 2151–2156.
[5] Inderfurth, K., Lindner, G., & Rachaniotis, N. P. (2005). Lot sizing in a production system with
rework and product deterioration. International Journal of Production Research, 43(7), 1355–
1374.

Page 29 of 29