Multiple Level Programming:

A Review

Hsu-Shih Shih, Ph.D.

Graduate Institute of Management Sciences

Tamkang University, Tamsui
Taipei, Taiwan, ROC
Personal Background
™ Ph.D., Industrial Engineering, Kansas State
Univ.
™ Visiting Professor, Univ. of Pittsburgh
™ Major Interests
¾ Decision analysis
¾ Decision support
¾ Operations research
¾ Soft computing
™ Website
¾ http://163.13.193.161
Brief Contents
™Introduction
™Definition
™Characteristics
™Applications
™Techniques
™Future Research
™Questions and Comments
Introduction
™ What is multiple level programming?
¾Decentralized planning in organizations
™ Where are its applications?
¾Many areas with conflict resolution
™ What’s techniques deal with the
problems?
¾Traditional and non-traditional techniques
™ Future Research
Definition
™ Multiple Level Programming (MLP)
¾To solve decentralized planning problems
with multiple executors in a hierarchical
organization
¾Explicitly assigns each agent a unique
objective and set of decision variables as
well as a set of common constraints that
affects all agents
Hierarchical Structure
MLP Formulation
Characteristics
™ Common Characteristics of MLP
1) Interactive decision-making units exit within a
predominantly hierarchical structure
2) Execution of decisions is sequential, from top
level to bottom level
3) Each unit independently maximizes its own net
benefits, but is affected by actions of other units
through externalities
4) The external effect on a decision-maker’s
problem can be reflected in both his objective
function and his set of feasible decision space
Bi-level Programming– a simple case
Problem formulation
Max f1 (x1, x2) = c11T x1 + c12T x2 (upper level)
x1

where x2 solves,
Max f2 (x1, x2) = c21T x1 + c22T x2 (lower level)
x2

s.t.
(x1, x2) ∈ X ={(x1, x2)| A1 x1+A2 x2≤ b, and x1, x2≥ 0}
where c11, c12, c21, c22, and b are vectors, A1 and
A2 are matrices, and X represents the constraint region.

Wen and Hsu 1991

Bi-level Programming
™ A Special Case of Two-person, Non-
zero Sum Non-cooperative Game
¾A general Stackelberg’s (leader-follower)
duopoly model
™ Nested Optimization Problem
¾NP-hard complexity
Applications (I)
™ Agricultural model
¾ Agricultural policy- Nile Valley case (Parraga 1981)
¾ Milk industry (Candler and Norton 1977)
¾ Mexican agriculture model (Candler and Norton 1977)
¾ Water supply model (Candler et al. 1981)
™ Government policy
¾ Distribution of government resources (Kyland 1975)
¾ Environmental regulation (Kolstad 1982)
™ Finance model
¾ Bank asset portfolio (Parraga 1981)
¾ Commission rate setting (Wen and Jiang 1988)
Applications (II)
™ Economic systems
¾ Distribution center problem (Fortuny and McCarl 1981)
¾ Principle-agent model (Arrow 1986)
¾ Price ceilings in the oil industry (DeSilva 1978)
™ Welfare
¾ Allocation model of strategic weapons (Bracken et al.
1977)
™ Transportation
¾ Highway network system (LeBlance and Boyce 1986)
™ Others
¾ Network flows (Shih and Lee 1999, Shih 2005)
¾ Supply chain (Viswanarthan et al. 2001)
Techniques (I)
™ Extreme-point Search
¾ Kth-best algorithm
¾ Grid-search algorithm
¾ Fuzzy approach (Shih 1995, Shih et al. 1996)
¾ Interactive approach (Shih 2002)
™ Transformation Approach
¾ Complement pivot
¾ Branch-and-bound
¾ Penalty function
™ Interior Point
¾ Primal-dual algorithm

Lee and Shih 2001, Shih et al. 2004

Techniques (II)
™ Decent and Heuristics
¾ Descent method
¾ Branch-and-bound
¾ Cutting plane
¾ Dynamic programming (Shih and Lee 2001, Shih 2005)

™ Intelligent Computation
¾ Tabu search
¾ Simulated annealing
¾ Genetic algorithm
¾ Artificial neural network (Shih et al. 2004)

Lee and Shih 2001, Shih et al. 2004

Categories of Techniques

Lee and Shih 2001,

Shih et al. 2004
Example 1. A Trade-off Problem
between Exports and Imports
Problem formulation
Max f1 = 2 x1 - x2 (effect on the export trade - 1st objective )
x1
where x2 solves,
Max f2 = x1 + 2 x2 (profits on the product - 2nd objective )
s.t.
3 x1 - 5 x2 ≤ 15 ( capacity )
3 x1 - x2 ≤ 21 ( management )
3 x1 + x2 ≤ 27 ( space )
3 x1 + 4 x2 ≤ 45 ( material )
x1 + 3 x2 ≤ 30 ( labor hours )
x1 , x2 ≥ 0 ( non-negative )

Shih 1995, Shih et al. 1996

Kth-best Algorithm– Extreme-point
™ Solving procedure
¾ Step 1. Solve the upper-level problem
i=1, x[1]*= (7.5,1.5) at vertex B
¾ Step 2. Solve the lower-level problem with x1 = 7.5
Solution x+= (7.5,4.5) between vertex D and vertex C
x+ ≠ x[1]*, go to Step 3.
¾ Step 3. Consider the neighboring set of x[1]* (vertex A and
vertex C)
¾ Step 4. Update label i=i+1=2, and choose x[2]* = (8,3)
(vertex C). Go to Step 2.
¾ Step 2. Let x1 = 8 to the lower level problem
Solution x+= (8,3). Since x+= x[2]*, the procedure is
terminated. x[2]* is the optimum
Decision (Variable) Space
Objective (Function) Space
Karush-Kuhn-Tucker Conditions–
Transformation approach
Problem formulation
Max f1 = 2 x1 - x2
x1, x2
s.t.
(x1 , x2 ) ∈ X
w1 (− 3 x1 + 5 x2 + 15) = 0
w2 (− 3 x1 + x2 + 21) = 0
w3 (− 3 x1 − x2 + 27) = 0
w4 (− 3 x1 − 4 x2 + 45) = 0
w5 (− x1 − 3 x2 + 30) = 0
− 5 w1 − w2 + w3 + 4 w4 + 3 w5 = 2
w1 , w2 , w3 , w4 , w5 , x1 , x2 ≥ 0

Lee and Shih 2001

Separation Procedure
Problem formulation
The constraint set
wT (A1 x1 + A2 x2 − b ) = 0, where w is a dual vector.

The transformed two terms

w ≤ (1 − η) M , and
A1 x1 + A2 x2 − b ≤ M η
where η ∈ {0, 1} and M is a large positive constant

Lee and Shih 2001

Concept of Fuzzy Approach
™ Fuzzy Membership Functions (Zadeh
1965)
¾ Tolerance of decisions
¾ Achievement of goal
™ Fuzzy Multi-objective Decision Making
(Zimmermann 1985)
¾ Information aggregation

Supervised search procedure

Shih 1995, Shih et al. 1996
Fuzzy Approach
Problem formulation
Max f2 = 2 x1 - x2
s.t.
(x1 , x2 ) ∈ X
μf1( f1(x)) ≥ α
μx1( x1) ≥ β
α ∈ [0, 1] and β ∈ [0, 1]
x1 , x2 ≥ 0

Shih 1995, Shih et al. 1996

Fuzzy Decision
Problem formulation
Max {α, β, δ}
s.t.
(x1 , x2 ) ∈ X
μf1( f1(x)) = (f1 − 0) / (13.5 − 0) ≥ α
μx1(x1) = (x1 − 4.5) / (7.5 − 4.5) ≥ β
μx1(x1) = (8 − x1 ) / (8 − 7.5) ≥ β
μf2( f2(x)) = (f2 − 10.5) / (21 − 10.5) ≥ δ
x 1 , x2 ≥ 0
α, β, δ ∈ [0, 1]

Shih 1995, Shih et al. 1996

Interactive Approach

Shih 2002
Advantages of Fuzzy Approach
™ Advantages
¾Approximation of the natural of Large
MLPPs
¾Not increase the computational complexity
¾Ease to extend to multiple levels
¾DMs involve the process
¾Efficient (Pareto) solution

Nested Optimization ⇒ Sequential Optimization

Extension to Vague Information

Vague/Imprecise data ⇒ Possibilistic Distribution

Shih 2002
Dynamic Aspect of MLP

Dynamic environment ⇒ Multi-stage MLP

(discrete space)
Shih 2005
Neural Network Approach
™ Use of dynamic behavior of artificial
neural networks with parallel processing
™ Based on Hopfield and Tank (1985)-
recurrent network
™ Transforming to the energy function
without constraints
™ Optimum solution with a steady state

Shih et al. 2004

Neural Network Approach
Future Research
™ Use of hybrid algorithms for uncertainty
™ Solutions of multi-subunits
™ Extend to n-level problems
™ Conditions of existing Pareto-optimal
™ Applications of real-world problems
(nonlinear coefficients)
Reference
Questions & Comments

Thank you!