pclsp ntnu
DESCRIPTION
Talking about Dynamic Pricing and Lot-sixing at NTNUTRANSCRIPT
The profit maximizing capacitated
lot-size (PCLSP) problem
by Kjetil K. Haugen
Asmund Olstad
and
B̊ard I. Pettersen
Molde University College
Servicebox 8, N-6405 Molde, Norway
E-mail: [email protected]
European Journal of Operational Research, 176:165–176, 2006
1
Idea – Abstract
• Introduce a ”new” set of lot-sizing models
including pricing – PCLSP.
• PCLSP – practically at least as relevant
as CLSP.
• PCLSP – computationally more feasible
than CLSP.
2
Outline
1) Introduce LSP, CLSP and PCLSP.
2) Brief discussion of algorithmic properties
of LSP and CLSP.
3) Introduce a Lagrange Relaxation algo-
rithm for PCLSP.
4) Judge algorithmic performance by exam-
ples and compare with CLSP
5) Discuss practical relevance of PCLSP.
3
The simple lot-size problem (LSP) (1)
Min Z =T∑
t=1
stδt + htIt + ctxt (1)
s.t.
xt + It−1 − It = dt ∀t (2)
0 ≤ xt ≤ Mtδt ∀t (3)
It ≥ 0, ∀t (4)
δt ∈ {0,1} ∀t (5)
4
The simple lot-size problem (LSP) (2)
Decision variables:
xt = amount produced in period t
It = inventory between t, t + 1
δt = 1 if xt > 0 in period t ; 0 otherwise
Parameters:
T = number of time periods
st = setup cost in period t
ht = storage cost between t, t + 1
ct = unit production cost in period t
Mt = ”Big M” in period t
5
The simple lot-size problem (LSP) (3)
– Problem characteristics
Basic trade off:
• Many set-ups ⇒ ∑t stδt ↑ and
∑t htIt ↓
• Few set-ups ⇒ ∑t stδt ↓ and
∑t htIt ↑
Slight generalization of EOQ-model:
Given dt = d, ct = c, ht = h∀t ⇒ LSP = EOQ
6
The simple lot-size problem (LSP) (4)
– Algorithmic characteristics
• Polynomial DP-algorithm by Wagner and
Whitin (1950).
• Solves very fast.
• Planning horizon theorems; xt · It = 0.
• Interesting candidate as sub-problem
solver in more advanced lot-size problems.
7
The capacitated problem (CLSP) (1)
Min Z =T∑
t=1
J∑
j=1
sjtδjt + hjtIjt + cjtxjt (6)
s.t.
J∑
j=1
ajtxjt ≤ Rt ∀t (7)
xjt + Ij,t−1 − Ijt = djt ∀jt (8)
0 ≤ xjt ≤ Mjtδjt ∀jt (9)
Ijt ≥ 0, ∀jt (10)
δjt ∈ {0,1} ∀jt (11)
8
CLSP (2)
Decision variables:
xjt = amount of item j produced in t
Ijt = inventory of item j between t, t + 1
δjt =
{1 if item j is produced in period t0 otherwise
Parameters:
T = number of time periods
J = number of items
sjt = setup cost for item j in period t
hjt = storage cost, item j between t, t + 1
cjt = unit production cost, item j at t
ajt = resource used, item j at t
Rt = capacity resource available at t
Mjt =T∑
s=t
djs
9
CLSP (2) – characteristics
• Much harder to solve compared to LSP.
• Reason:
Violation of capacity constraint ⇒”moving production around” (combi-natorial).
• Due to non existence of polynomial al-gorithms (NP-hardness) heuristical (La-grange relaxation based) approaches com-mon.
• A very ”popular” OR research problem.
• Still: typical problem sizes not muchlarger than 100 x 100 – not satisfactorygiven product variety today.
10
PCLSP (1)
Max Z =T∑
t=1
J∑
j=1
[djtpjt − sjtδjt − hjtIjt − cjtxjt
]
(12)
s.t.
αjt − βjt · pjt = djt ∀jt (13)J∑
j=1
ajtxjt ≤ Rt ∀t (14)
xjt + Ij,t−1 − Ijt = djt ∀jt (15)
0 ≤ xjt ≤ Mjtδjt ∀jt (16)
Ijt ≥ 0, ∀jt (17)
δjt ∈ {0,1} ∀jt (18)αjt
βjt≥ pjt ≥ 0 ∀jt (19)
11
PCLSP (2)
Decision variables added to CLSP:
pjt = price of item j in period t
Parameters added to CLSP:
αjt = constant in linear demand,
item j at t
βjt = slope in linear demand,
item j at t
12
PCLSP characteristics
• Linear demand, Monopoly assumption(unrealistic).
• PCLSP is a generalization of CLSP. ∗
• Immediate feasible solutions are obtain-able (as opposed to CLSP) by ”pricingout”. †
• PLSP (uncapacitated single item ver-sion) is well known from OR-literature.Thomas (1970) constructed a polynomialDP-algorithm with complexity as of theWagner/Whitin algorithm.
∗Easy to see by the special case pjt = p̂jt where p̂jt areassumed constant.†That is, any capacity constraint violation can be”removed” as any demand may be forced to zero(pjt = αjt
βjt).
13
Basic hypothesis
The added problem flexibility ob-
tained by introduction of price vari-
ables should make the problem easier
to solve.
14
LUBP sub problems
By relaxing the capacity constraint (14), the
PCLSP-problem (12) – (19) may be expressed
as:
Max Z = Z +T∑
t=1
λt
Rt −
J∑
j=1
ajtxjt
s.t. constraints (13) to (19)
It is ”straightforward” to adjust the DP-
algorithm by Thomas (1970) to provide very
efficient solutions to the LUBP sub-problems.
Note: Solving LUBP yields upper bound on
PCLSP.
15
LLBP sub problems
If δij’s are fixed (Set-up structure) in PCLSP,
a standard quadratic programming problem is
obtained:
Max Z =T∑
t=1
J∑
j=1
[djtpjt − hjtIjt − cjtxjt
]
s.t. (13), (15), (17), (19).
Any solution to LLBP is feasible and typically
non-optimal. Hence, Any solution to LLBP
is a lower bound to PCLSP.
16
Algorithmic structure
LUBP
LLBP
Z*
PCLSP
Set-upstructure λλλλ
t
k
Algorithmic structure
17
Algorithm: Lagrange relaxation
0. Define λt = 0,∀t = 1,2, . . . , T
1. Solve LUBP (Obtain Set-up structure)
2. Solve LLBP (for Set-up structure ob-tained in step 1.) Define λt from thissolution as λk
t , where k denotes iterationcount.
Stop if: (define reasonable stopping crite-ria)
3. Update λkt by smoothing: λk+1
t = θk ·λk−1
t +(1−θk)λkt ,∀t, where θk is a smooth-
ing parameter, 0 ≤ θk ≤ 1
4. Go to step 1.
18
Some results (1)
CLSP PCLSPProblem #It. G(%) #It. G(%)CAP93% 50 10 17 0.59CAP91% 50 3 32 0.41CAP73% 50 2 6 0.2CAP61% 50 1 22 0.06
• Cases captured from Thizy and Wassen-
hove (1985). 8 products over 8 time pe-
riods.
• #It. means number of iterations per-
formed in algorithm.
• G(%) means gap in percent orZLUBP−ZLLBP
ZLUBP· 100.
19
Some results (2)
CLSP PCLSPProblem #It. G(%) #It. G(%)CAP93% 50 10 1 3.68CAP91% 50 3 1 2.82CAP73% 50 2 1 1.5CAP61% 50 1 1 0.65
The same cases with Gaps after one itera-
tion in PCLSP compared to 50 iterations in
CLSP.
20
PCLSP – practical relevance
Demonstrated (by some examples) that
PCLSP solves significantly faster than CLSP.
What about practical relevance?
Obvious facts:
• If monopolistic market conditions ⇒PCLSP Â CLSP
• If Free markets (price taking behavior)
CLSP is ”correct”
• Most markets are neither (oligopoly) –
what then?
21
Price constraints
• Given ”relatively small” price changes, un-
derlying Nash equilibrium may be stable.
• Price constraints may serve the purpose
• Relatively simple to introduce (no radical
algorithmic changes)
22