Department of Systems Engineering and Operations Research

SYST 542
Decision Support Systems
Engineering
Instructor: Kathryn Blackmond Laskey
Spring Semester, 2006

Unit 3: DSS Elements:

The Model Subsystem (1)
Decision Analysis & Optimization
SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 1 -
 Department of Systems Engineering and Operations Research

Outline

• Developing the model subsystem

• Role of decision theory in DSS
• Brief survey of decision analysis
and optimization methods

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 2 -

 Department of Systems Engineering and Operations Research

Models and DSS

• A model is a representation of a system which
can be used to answer questions about the
system
• A DSS uses computer models in conjunction
with human judgment
– Performs computations that assist user with decision problem
– Design is based on a model of how human user does / ought to solve
decision problem

• Model subsystem can be:

– completely automated
– partially automated
– manual with automated support for information entry,
retrieval and display

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 3 -

 Department of Systems Engineering and Operations Research

DSS and Exploratory Models

• DSS modeling is by definition exploratory
– Human remains in the loop
• Consolidative model may be possible for parts
of problem
– Avoid the temptation to pour too many resources into the
part you know how to model!
• Good DSS helps DM make use of partial
information
– to generate hypotheses about system behavior
– to demonstrate occurrence of types of behavior under not-
too-implausible assumptions
– to explore possible risks / failure modes
– to determine regions of parameter space in which certain
qualitative behaviors occur

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 4 -

 Department of Systems Engineering and Operations Research

Issues for Exploratory Modeling

• Representing the ensemble of models
– internal system representation
– decision maker’s mental model
– language for communicating with decision maker
• Tools for allowing DM to explore alternative
modeling assumptions
– what-if analysis
– sensitivity analysis
– exploring different parts of parameter space
– exploring different combinations of modeling assumptions
• Techniques for helping DM assess consequences of
alternative assumptions
– summaries of high-dimensional data
– graphical displays
SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 5 -
 Department of Systems Engineering and Operations Research

Steps in Developing
the Model Subsystem
1. Map functions in decision process onto
models
2. Determine input / output requirements for
models
3. Develop interface specifications for models
with each other and with dialog and data
subsystems
this step may result in additional modeling activity
4. Obtain / develop software realizations of the
models and interfaces

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 6 -

 Department of Systems Engineering and Operations Research

Decision Process (review)

Identify Problem

Identify Objectives
(values)

{
Identify Alternatives

Decompose and Model Problem

– Structure
Use of – Uncertainty
DSS – Preference

Choose Best Alternative

Sensitivity Analysis

More
Analysis Yes
Needed
No
Make Recommendation Thanks to Andy Loerch

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 7 -

 Department of Systems Engineering and Operations Research

Some Typical Problems to Model

• Evaluate benefits of proposed policy against costs
• Forecast value of variable at some time in the
future
• Evaluate whether likely return justifies investment
• Decide where to locate a facility
• Decide how many people to hire & where to assign
them
• Plan activities and resources for a project
• Develop repair, replacement & maintenance policy
• Develop inventory control policy

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 8 -

 Department of Systems Engineering and Operations Research

A Brief Tour of Modeling Options

• A wide variety of modeling approaches is
available
• DSS developer must be familiar with broad
array of methods
• It is important to know the class of problems
for which each method is appropriate
• It is important to know the limitations of each
method
• It is important to know the limitations of your
knowledge and when to call in an expert

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 9 -

 Department of Systems Engineering and Operations Research

Decision Theory
• Formal theory to support GOOD-D process
• Goals (What do I want?)
– Begin with value-focused thinking
– Quantify values with utility function
• Options (What can I do?)
• Outcomes (What might happen?)
– Quantify uncertainty with probability distribution
• Decide:
– Develop a mathematical model of expected utility for each option
– Model recommends the option for which expected utility is greatest
– In a good decision analysis, model building process increases
understanding of decision problem
– The model gives insight but the decision maker makes the final choice
• Do it!
– Discussion and evaluation of options should consider issues of
implementation

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 10 -

 Department of Systems Engineering and Operations Research

Role of Decision Theory in DSS

• Avoid “elicit model out of decision maker’s
head, push the button and solve for the correct
answer” mentality
• Decision theoretic models are appropriate
when:
– We can quantify values and uncertainties to a reasonable
approximation
– It is useful to suggest potentially optimal solutions and/or
to weed out clearly suboptimal solutions
• Useful outputs (in addition to recommended
solution)
– Explanation of results
– Sensitivity analysis
– Visualization of feasible region

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 11 -

 Department of Systems Engineering and Operations Research

Decision Analysis
• Collection of analytic and heuristic procedures for
developing decision theoretic model
• Goals of decision analysis
– Organize or structure complex problems for analysis
– Deal with tradeoffs between multiple objectives
– Identify and quantify sources of uncertainty
– Incorporate subjective judgments
• Decision analysis methods help to:
– decompose problem into subproblems which are easier to solve
– detect and resolve inconsistencies in solutions to the
subproblems
– aggregate solutions to subproblems into a consistent action
recommendation for the original problem

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 12 -

 Department of Systems Engineering and Operations Research

Decision Analysis Methods

• Value Models: Multiattribute Utility
• Uncertainty Models: Decision Trees
– A structured representation for options and outcomes
– A computational architecture for solving for expected utility
– Best with “asymmetric” problems (different actions lead to qualitatively
different worlds)
• Uncertainty Models: Influence Diagrams
– A structured representation for options, outcomes and values
– A computational architecture for solving for expected utility
– Best with “symmetric” problems (different actions lead to worlds with
qualitatively similar structure)
• Decision analysis software:
– http://faculty.fuqua.duke.edu/daweb/dasw.htm (there are some broken
links)

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 13 -

 Department of Systems Engineering and Operations Research

Example: Patient Treatment

A patient is suspected of having a disease. Treated patients Goals:
recover quickly from the illness, but the treatment has
• Recovery
unpleasant side effects. Untreated patients suffer a long
• Freedom from side effects
and difficult illness but eventually recover.
Options:
• Treat of don’t treat
Utility
Outcomes:
• Sick/Well
• Side Effects / No Side Effects
Speed of Side
Recovery Effects
Multiattribute
Hierarchy UT
Treat

Disease UD
Don’t
Disease treat No disease
UN
Outcome Utility
Decision Tree

Treatment
Influence Diagram

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 14 -

 Department of Systems Engineering and Operations Research

Value Model

• Objectives related to alternatives by Attributes

• Attributes are measures of achievement of objectives
– Quantitative
– Reflect consequences
• Usually decision maker has multiple objectives
– Objectives are often in conflict
– Value model incorporates tradeoffs among objectives
• Types of value model
– Ordinal - ranking only
– Measurable value function - strength of preference
– Utility function - includes risk attitude
• Medical example:
– Need to assess relative degree of misery of side effects vs illness
– Need utility model to trade off chance of illness against cost of
side effects

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 15 -

 Department of Systems Engineering and Operations Research

Constructing a Value Model

• Decompose objectives
– Independent components of value (avoid double-counting)
– Begin with fundamental objective and decompose into important means
objectives
• Find ways to measure objectives
– Natural attribute (e.g., cost in dollars, weight in pounds)
– Constructed attribute (e.g., consumer price index for inflation)
– Proxy attribute (e.g., sulfur dioxide emissions for erosion of monuments
from acid rain)
• Combine objectives
– Turn attribute scores into value function
» Better options have higher value
» Equal differences in value function are equally valued by DM
– Functional form depends on relationship between attributes
» Most common combination method is linear additive with cutoffs
» Justification depends on independence assumptions
– Weights trade off objectives against each other
» Subjective
» Need to consider range of weights
• Adjust for risk attitude if necessary

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 16 -

 Department of Systems Engineering and Operations Research

Linear Additive Value Function

• Value function is weighted sum of single-
attribute value functions
– v(x1, …, xn) = w1v1(x1) + … + wnvn(xn)
• Requires attributes to be preferentially
independent:
– Preference order between levels of any pair Xi and Xj of
attributes does not depend on levels of other attributes
• Much simpler to specify and use than more
complex functional forms
• Try to specify attributes to be preferentially
independent

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 17 -

 Department of Systems Engineering and Operations Research

Example
Multiattribute Hierarchy:
Buying a Beach House
• Decompose value into attributes
Initial – nonoverlapping
Investment – cover all important aspects of value
Financial – bottom level attributes are measurable
• Assess function for combining
NPV attributes at each level (usually linear
weighted average)
Total
Utility • Compute utilities of all options
– score on bottom-level attributes
Time Spent – compute overall score

Enjoyment Luxury
Walking time
Ocean access
View

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 18 -

 Department of Systems Engineering and Operations Research

Assessing Weights:
Swing Weight Method
• First weight
– Imagine all attributes are at worst level (may be imaginary)
– Which would you choose to increase to best level?
– Assign this attribute weight of 1
• Rest of weights
– All attributes are at worst level again Beware: Some commonly
– Pick another attribute to move to best level used weight assessment
– What % of value of moving first to its best level? methods ignore absolute
• Scale all weights to sum to 1 scale of attributes and can
lead to preference reversals.

Best

70% Best w2 = 0.7 w1

w1 + w2 = 1

w1 = 0.59
w2 = 0.41
Worst Worst

Attribute 1 Attribute 2

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 19 -

 Department of Systems Engineering and Operations Research

Analytic Hierarchy Process

• Popular method for building a preference
model
• Problem decomposition into multiattribute
hierarchy is same as for multiattribute utility
• Method of assigning weights is different
– Based on paired comparisons
– Pairs of options are compared on scale from 0 to 9
– Ratings are used to develop weights for the value function
• Comments
– Method is popular because paired comparisons are natural
and intuitive to many decision makers
– Theoretical justification of the MAU “swing weight”
assessment is lacking
– Can have preference reversals when options are added or
removed from the option set (i.e., whether we prefer A to B
may depend on whether or not C is under consideration)

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 20 -

 Department of Systems Engineering and Operations Research

Decision Analysis Example:

Texaco vs Pennzoil (1984)
• Pennzoil and Getty agreed to merge
• Texaco made Getty a better offer - Getty reneges
• Pennzoil sues, wins case in 1985, get $11.1 Billion
• Texas appeals court reduces judgment by $2 Billion
– With court costs and interest $10.3 Billion
– Texaco threatened to bankrupt and go to Supreme Court
• 1987, before Pennzoil starts issuing liens Texaco offers to settle
for $2 Billion
• Pennzoil thinks $3-5 Billion is a fair price
• What should Hugh Liedtke, CEO of Pennzoil, do?

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 21 -

 Department of Systems Engineering and Operations Research

Decision Tree for Pennzoil’s Problem

(simplified model)
Result ($B)
Accept $2 Billion
2
Texaco Accepts $5 Billion
5
0.2
0.17 10.3
Counteroffer Final Court 0.5
5
$5 Billion 0.50 Decision
Texaco 0.3
Refuses 0
Counteroffer 0.2
10.3
Texaco Final Court 0.5
Counteroffers 0.33 5
Decision
$3 Billion 0.3
0

Accept $3 Billion 3
How could this model be made more complex?

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 22 -

 Department of Systems Engineering and Operations Research

Influence Diagram
• Alternative representation of
Activity Carcinogenic
decision problem
Test Activity – Ovals are “chance nodes”
– Boxes are “decision nodes”
– Rounded boxes are “value nodes”
Exposure Human
Test Exposure – Arcs show influences
• Formally equivalent to decision
Cancer
tree
Cost – Probability and utility values are
Usage Net encapsulated inside the nodes
Decision Value – Some software packages switch back
and forth between views

Economic
Value

• Dotted lines are information arcs

• Whether to collect information can be represented as a decision problem
• Note: influence diagram represents multiattribute utility function explicitly

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 23 -

 Department of Systems Engineering and Operations Research

Some Simple
Qualitative Rules
• Dominance
– If Option X is at least as good as Option Y on all attributes
of value, Option X is at least as good as Option Y
– If Option X is at least as good as Option Y for each
possible outcome, then Option X is at least as good as
Option Y
• Useless Information: If information gathering
is costly and the result would not change your
decision, then do not gather the information

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 24 -

 Department of Systems Engineering and Operations Research

Mathematical Programming
• Constrained optimization problems:
– Maximize or minimize objective function
– Subject to constraints defining feasible region of solution space
• Solution methods:
– Linear programming (LP)
» Objective function and constraints are linear
– Nonlinear programming (NLP)
» Objective function and/or some constraints are nonlinear
– Integer programming (IP)
» Feasible space consists of integer variables
– Mixed integer programming (MIP)
» Feasible space consists of some integer and some real variables
– Goal programming (GP)
» Try to find at least one solution in feasible region
– Dynamic programming (DP)
» Find optimal policy in sequential decision making problem
• Traditional mathematical programming ignores uncertainty
SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 25 -
 Department of Systems Engineering and Operations Research

LP Example
• A company makes 3 types of furniture:

Type Profit Labor Materials Minimum

/item Required Required Qty
(hours) (sq ft)
Chair $50 10.5 5 5
Bench $100 15 15 7
Table $75 17 10 5

° Objective: Find the highest profit combination of items to manufacture

° Constraints:
- Labor hours available = 400
- Lumber available = 300
- Must make at least minimum quantity of each item

Thanks to Andy Loerch

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 26 -

 Department of Systems Engineering and Operations Research

LP Formulation
Maximize 50 c + 100 b + 75 t profit

s.t. 10.5 c + 100 b + 17 t ≤ 400 labor

5 c + 15 b + 10 t ≤ 300 lumber
c ≥ 5 chairs
b ≥ 7 benches
t ≥ 5 tables

Thanks to Andy Loerch

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 27 -

 Department of Systems Engineering and Operations Research

Solving Linear Programs

• Simplex method - developed by Dantzig in 1940’s
– Standard method
– Exponential in number of variables
– Guaranteed to give optimal solution
– Searches extreme points in feasible region
• Karmarkar’s algorithm - 1980’s
– Polynomial time
– Very fast on large problems
– Limited ability to do sensitivity analysis
• Specialty algorithms exploit special case
structures
– Transportation method
– Network simplex

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 28 -

 Department of Systems Engineering and Operations Research

Goal Programming
• Define goals (aspiration levels) as constraints:
– f(x) ≥ b; f(x) ≤ b; f(x) = b
• In standard LP these would be constraints
defining feasible region
• In GP we try to minimize deviation from goal
– Minimize weighted sum of goal deviations
– Minimize some other function of goal deviations
– Minimize worst deviation
– Lexicographically minimize ordered set of goal deviations

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 29 -

 Department of Systems Engineering and Operations Research

Solving Integer Programs

• Most IPs and MIPs are binary
– General integers expressed as sums of binaries with
rounding
• Standard method: Branch and bound
– Solve LP with integer constraints relaxed
– Choose a variable to branch on
» Make 2 problems - set chosen variable to 1 or 0
» Solve both relaxed problems
– Repeat till best integer solution is found
– Worst case: 2n LPs to solve
» Can explode rapidly

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 30 -

 Department of Systems Engineering and Operations Research

Solving Nonlinear Programs

• Standard methods
– Steepest descent
– Conjugate gradient
• Convexity is important
– Using standard NLP solvers on non-convex problems can
give local (not global) optimum!!
– Stay tuned (next week) for more on non-convex problems!

Non-convex Function
g(x)

Local min
Global min
x* x
SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 31 -
 Department of Systems Engineering and Operations Research

Solving Mathematical Programs

• Special purpose optimization packages
– e.g., OSL, CPLEX
– Linear, nonlinear, integer programs
• Spreadsheet add-ins
– e.g., Excel’s solver
– Easily available, don’t need to learn new package or
interface to external software
– Usually limited (e.g., LP only; size limits)
• Many problems cannot be solved exactly
– Heuristic methods are used
– Interface between AI and OR/MS

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 32 -

 Department of Systems Engineering and Operations Research

Solving LP Using Excel Solver

(1) Logically organize data (label, etc.)
• Coefficients for objective function
• Coefficients for constraints
• RHS of the constraints
(2) Reserve cells for the decision variables
– Called Changing Cells
(3) Create formula in a cell for the objective function
– Called Target Cell
(4) Create a formula for the LHS of each constraint
(5) Open Solver Dialog box (Tools menu)
(6) Enter the appropriate info and run Solver

Thanks to Andy Loerch

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 33 -

 Department of Systems Engineering and Operations Research

Sensitivity Analysis
• One-variable sensitivity analysis
– How sensitive is solution to change in parameter (weight in
objective function or constraint value)?
– Simplex method can produce one-variable sensitivity
analysis as a by-product
• Parametric analysis
– Specify range of values for parameter or parameters
(weight on objective function; value of constraint;
probability)
– Evaluate change in solution as parameters vary through
range

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 34 -

 Department of Systems Engineering and Operations Research

Visualizing Sensitivity
Analysis Results
Sensitivities to Parameters
• Tornado Diagram
D
– Visualizes result of varying a L
set of parameter through
R
specified ranges on an
output of interest S

• Strategy Region Graph

Parameter 1
– Visualizes changes in
optimal strategy as 2
parameters are varied
through a range

Parameter 2

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 35 -

 Department of Systems Engineering and Operations Research

In Summary...

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 36 -

 Department of Systems Engineering and Operations Research

References
• Anderson, D., Williams, T., and Sweeney, T., An Introduction to Management
Science: Quantitative Approaches to Decision Making, Southwestern, 1999.
• Clemen, R. Making Hard Decisions: An Introduction to Decision Analysis,
Duxbury, 1997.
• Winston, W. Operations Research Applications and Algorithms, Duxbury, 1997.

SYST 542 Copyright © 2006, Kathryn Blackmond Laskey Unit 3 - 37 -