Hidden (Tabu) Secrets of Successful Evolutionary Search
Methods
Peng

Yeng Yin
National Chi Nan University, Taiwan
Fred Glover
University of Colorado, Boulder, CO, USA
2007 IEEE Congress on Evolutionary Computation
P.Y. Yin & Fred Glover
2
New Directions for Evolutionary Algorithms
•
We are witnessing a radical departure from the
early 1990s
•
Original theme for Solution Combination:
restrict attention to a handful of genetically
inspired “crossover” operations. (Basic simple
approaches suffice.)
•
Today
emphasize
:
–
New
solution
codings
–
New
forms
of
crossover
–
Exploit
problem

specific
knowledge
P.Y. Yin & Fred Glover
3
Limitations Still Exist!
Regarding Solution Combinations and Their Use
We still often focus on narrow metaphors for
combining solutions
•
Advantages
of
the
metaphors
:
–
Stimulate
researchers
to
create
methods
based
on
the
metaphors
–
Catch
the
attention
of
wide
audiences
(Easy
to
describe
.
Great
sources
of
publicity!)
•
Disadvantages
of
the
metaphors
:
–
Create
a
biased
perspective
about
preferred
forms
of
methods
–
Cause
valuable
alternative
strategies
to
be
overlooked
P.Y. Yin & Fred Glover
4
Remedy
–
An Orientation Shift
•
View Solution Combinations as a process of
Path
Relinking
•
Employ frameworks of
Adaptive Structured
Combinations
•
Exploit the
neighborhood concept
in generating
combinations
•
Integrate with
adaptive memory
P.Y. Yin & Fred Glover
5
Structured Combinations for Populations of Vectors
•
Property 1 (Representation Property). Each vector
represents a set of votes for particular decisions
•
Property 2 (Trial Solution Property). The set of votes
translates into a trial solution by a well

defined process
•
Property 3 (Update Property). If a decision is made
according to the votes of a given vector, thus producing
a residual problem requiring fewer decisions, a well

defined rule exists to update all vectors relative to the
residual problem so that Properties 1 and 2 continue to
hold
(Details and illustrations in F. Glover (1994) “Tabu
search for nonlinear and parametric optimization (with
links to genetic algorithms),”
Discrete Applied
Mathematics
49, 221

235.)
P.Y. Yin & Fred Glover
6
Path Relinking
X
A
B
C
Using A and
B as guiding
solutions
Using B and
C as guiding
solutions
Using C
as a
guiding
solution
X is the initial
solution
•
Valuable
information
can
be
contained
in
trajectories
from
bad
solutions
to
good
solutions
(or
from
good
solutions
to
other
good
solutions)
•
Path
relinking
selects
moves
that
introduce
attributes
contained
in
the
guiding
solution(s)
into
the
initial
solution
.
Basically
only
one
attribute
of
the
initial
solution
is
modified
at
each
time
a
move
is
made
.
•
Instead
of
interchanging
information
between
solutions
in
a
wholesale
fashion,
path
relinking
is
a
stepwise
approach
considering
neighborhood
structure
and
adaptive
memory
strategies
that
may
generate
solutions
not
accessible
by
EAs
P.Y. Yin & Fred Glover
7
Scatter Search
•
Useful
information
about
the
form
or
location
of
the
global
solution
is
typically
contained
in
a
sufficiently
diverse
collection
of
elite
solutions
Repeat until 
P
 =
PSize
P
Diversification Generation
Method
Subset Generation
Method
Improvement
Method
Solution Combination Method
Improvement
Method
No more new
solutions
Reference Set
Update Method
RefSet
Diversification Generation
Method
Improvement
Method
Stop if
MaxIter
reached
P.Y. Yin & Fred Glover
8
Neighborhood Construction
•
Candidate
list
strategy
–
Effectively
construct
a
neighborhood
structure
to
save
computational
time
or
even
to
construct
a
more
complex
(or
compound)
neighborhood
that
is
not
accessible
using
simple
moves
.
It
accommodates
variable
neighborhood
search
(VNS)
method
Ejection
chain
Filter
and
fan
P.Y. Yin & Fred Glover
9
Ejection Chain
•
Construct
neighborhood
by
combining
successive
interindependent
(component)
moves
to
form
a
single
compound
move
.
•
Tunnel
infeasible
region
by
successive
ejection
moves
and
transform
to
a
feasible
solution
by
a
trial
move
.
•
This
form
of
compound
neighborhood
structures
are
usually
not
accessible
by
feasibility

preserving
search
methods
.
t
tip
r
s
2
s
1
root
subroot
subroot
t
tip
r
s
2
s
1
root
subroot
subroot
Ejection move
Trial move
Traveling salesman problem
P.Y. Yin & Fred Glover
10
Filter and Fan
•
The neighborhood tree is explored
breadth first and is restricted by a
maximum number of levels L. Each
level is governed by the
filter
strategy
that selects a subset of moves induced
by the
fan
strategy
•
Filter & Fan is usually applied after a
local heuristic has been executed to
explore a larger neighborhood in
order to overcome local optimality
•
The two search strategies are
alternated when a new local optimum
is found until the Filter & Fan fails to
improve the current best solution
Protein
folding
problem
P.Y. Yin & Fred Glover
11
Strategic Oscillation
•
In many cases the elite (or globally
optimal) solution lies on the feasibility
boundary or the search method would
stop at a critical level
•
Strategic oscillation drives the search
toward or away from an
oscillation
boundary. The approach proceeds for a
specified depth beyond the boundary, and
turns around. The boundary again is
approached and crossed from the opposite
direction.
•
The oscillatory behavior is established by
generating modified evaluations and rules
of movement, depending on the regional
locality and trajectory direction
•
Choose oscillation pattern and change
rate to achieve an effective interplay
between intensification and
diversification
Oscillation boundary
Depth
Iterations
P.Y. Yin & Fred Glover
12
Adaptive Memory Programming
•
Adaptive memory programming (Tabu Search)
constitutes of adaptive memory and a set of
responsive strategies
Adaptive memory:
purposefully comparing previous
states or transactions to those currently contemplated
Recency, frequency, influence, quality
Responsive strategies: take advantage of adaptive
memory to exploit good solution features while
explore new promising regions
Tabu restriction, aspiration criteria
Intensification, diversification
P.Y. Yin & Fred Glover
13
Adaptive Memory Programming
•
Adaptive memory programming (Tabu Search)
Tabu restriction
: In each move iteration, the best
move (evaluated in aspects of quality, influence,
frequency) in the neighborhood is selected unless it is
tabu.
Tabu tenure
(static/dynamic/reactive) is
determined according to the adaptive memory
Aspiration criteria
: The tabu restriction can be
overruled if the corresponding move meets the
aspiration criteria such that the search can be guided
to a promising region along the course
Intensification/diversification strategies
: applying
incentives/penalties to induce attributes of good
solutions
path relinking, strategic oscillation, multi

start
P.Y. Yin & Fred Glover
14
Create Hybrid
–
Integrate EA and TS
Solution
combination
Path
relinking
exploiting
neighborhood
structure
and
adaptive
memory
–
directly
address
problem
context
without
metaphor
restrictions
–
stepwise
and
systematic,
reducing
the
probability
of
missing
valuable
information
on
the
course
Solution
improvement
Adaptive
memory
strategies
can
be
directly
applied
to
improve
the
combined
solution
–
accommodate
implicit
variable
neighborhood
search
that
prevents
from
local
optimality
Population
distribution
control
Tabu
search
has
provided
a
wealth
of
diversity
control
strategies
–
tabu
restrictions,
altering
move
evaluations,
strategic
oscillation
Constraint
handling
Strategic
oscillation
repeatedly
drives
the
search
across
the
feasibility
boundary
–
non

monotonically
change
the
mixes
of
feasible
and
infeasible
solutions
P.Y. Yin & Fred Glover
15
For Additional Background
Using
Google
search
:
“Path
Relinking”
returns
about
28
,
000
web
pages
.
“Scatter
Search”
returns
about
50
,
000
web
pages
.
The
first
references
encountered
on
Google
give
a
good
background
for
basic
understanding
.
(Incidental
remark
:
“Tabu
Search”
returns
about
530
,
000
web
pages
.
)
P.Y. Yin & Fred Glover
16
•
Portfolio
Management
•
Supply
Chain
Applications
•
Strategic
and
Operational
Planning
•
Financial
Planning
•
Manufacturing
Process
Flow
•
Resource

Constrained
Scheduling
•
Network
Planning
•
Routing
&
Distribution
•
Data
Mining
•
Biotechnology
•
Health
Care
OptTek Customized Simulation Optimization Applications
P.Y. Yin & Fred Glover
17
Metaheuristic
–
Based Simulation Optimization
P.Y. Yin & Fred Glover
18
•
Function
to
be
Optimized
•
Highly
Nonlinear
•
Nondifferentiable
•
Discrete or Continuous or Mixed
•
Function
Evaluations
•
Complex
•
Extremely
Computation
Intensive
•
One
second
to
One
Day
per
Evaluation!
The Optimization Challenge
P.Y. Yin & Fred Glover
19
•
Evolutionary
Scatter
Search
•
Advanced
Tabu
Search
•
Linear
&
Mixed
Integer
Programming
•
Pattern
Classification
&
Curve
Fitting
•
Neural
Networks
•
Support
Vector
Machines
&
Trees
•
SAT
Data
Mining
OptQuest
®
Components
P.Y. Yin & Fred Glover
20
Efficiency is Critical!
OptQuest
®
vs. RiskOptimizer
P.Y. Yin & Fred Glover
21
•
Given a set of opportunities and limited
resources…
•
…determine the best set of projects that
maximizes performance
Problem
P.Y. Yin & Fred Glover
22
•
Constraints
:
•
Budget
•
Resource
Availability
•
Scheduling
and
Sequencing
of
Projects
•
Project
Dependencies,
etc
.
•
Objectives
:
•
Maximize
N
et
P
resent
V
alue
(NPV)
•
Maximize
I
nternal
R
ate
of
R
eturn
(IRR)
•
Maximize
B
usiness

C
ase
V
alue
(BCV)
Portfolio Selection Problem
P.Y. Yin & Fred Glover
23
•
5
Projects
:
•
T
ight
G
as
P
lay
Scenario
(TGP)
•
O
il
–
W
ater
F
lood
Prospect
(OWF)
•
D
ependent
L
ayer
Gas
Play
Scenario
(DL)
•
O
il
–
O
ffshore
P
rospect
(OOP)
•
O
il
–
H
orizontal
W
ell
Prospect
(OHW)
•
Ten
year
models
that
incorporate
multiple
types
of
uncertainty
•
Evaluation
Time
:
1
s
/
Scenario
Application Example
P.Y. Yin & Fred Glover
24
Determine
project
participation
levels
[
0
,
1
]
that
•
Maximize
E(NPV)
•
Keep
s
<
10
,
000
M
$
(Risk
Control)
•
All
projects
start
in
year
1
Base Case
TGP = 0.4, OWF = 0.4, DL = 0.8, OHW = 1.0
E(NPV) = 37,393
s
=9,501
Base Case
P.Y. Yin & Fred Glover
25
Determine
project
participation
levels
[
0
,
1
]
AND
starting
times
for
each
project
that
•
Maximize
E(NPV)
•
Keep
s
<
10
,
000
M
$
(Risk
Control)
•
Projects
may
start
in
year
1
,
2
,
or
3
TGP
1
= 0.6, DL
1
=0.4, OHW
3
=0.2
E(NPV) = 47,455
s
=9,513 10
th
Pc.=36,096
Deferment Case
Deferment Case
P.Y. Yin & Fred Glover
26
Determine
project
participation
levels
AND
starting
times
for
each
project
that
•
Maximize
P(NPV
>
47
,
455
M
$
)
•
Keep
10
th
Percentile
of
NPV
>
36
,
096
M
$
•
Projects
may
start
in
year
1
,
2
,
or
3
TGP
1
= 1.0, OWF
1
=1.0, DL
1
=1.0, OHW
3
=0.2
E(NPV) = 83,972
s
=18,522
P(NPV > 47,455) =
0.99
10
th
Pc.=43,359
Probability of Success Case
Probability of Success Case
P.Y. Yin & Fred Glover
27
•
Cash
Flow
Control
•
Capital
Expenditure
Control
•
Reserve
Replacement
Goals
•
Production
Goals
•
Finding
Costs
Control
•
Dry
Hole
Expectations
Control
•
Reserve
Goals
•
Net
Profit
Goals
Extensions…
P.Y. Yin & Fred Glover
28
Treatment
Patient Arrival
Emergency Room (ER)
Approach=
optimize
current process,
redesign process and
re

optimize
.
Release
Admit
Joseph DeFee, CACI, Inc.
Hospital Emergency Room Process
P.Y. Yin & Fred Glover
29
•
Nurses
•
Physicians
•
P
atient
C
are
T
echnicians
(PCTs)
•
Administrative
Clerks
•
E
mergency
R
ooms
(ER)
ER Resources
P.Y. Yin & Fred Glover
30
•
Minimize
E[
T
otal
A
sset
C
ost]
•
Subject
to
:
–
E[
C
ycle
T
ime]
for
Level
1
Patients
<
2
.
4
hours
–
Number
of
Nurses
between
1
and
7
–
Number
of
Physicians
between
1
and
3
–
Number
of
PCTs
between
1
and
4
–
Number
of
Clerks
between
1
and
4
–
Number
of
ER
between
1
and
20
Problem
P.Y. Yin & Fred Glover
31
•
Set
up
OptQuest
to
run
for
100
iterations
and
5
runs
per
iteration
•
Each
run
simulates
5
days
of
ER
operation
•
Results
:
–
Best
solution
found
in
6
minutes
–
E[TAC]
=
$
25
.
2
K
(
31
%
improvement
)
–
E[CT]
for
P
1
=
2
.
17
hours
Solution
P.Y. Yin & Fred Glover
32
Possible
to
improve
E[CT]
for
P
1
even
further?
Arrive at
ER
Transfer to
room
Receive
treatment
Fill out
registration
OK?
Released
Admitted
Into
Hospital
Y
N
Current Process
Arrive at
ER
Transfer to
room
Receive
treatment
Fill out
registration
OK?
Released
Admitted
Into
Hospital
Y
N
Redesigned Process
Process Redesign
P.Y. Yin & Fred Glover
33
•
Set
up
OptQuest
to
run
for
100
iterations
and
5
runs
per
iteration
•
Each
run
simulates
5
days
of
ER
operation
•
Results
:
–
Best
solution
found
in
8
minutes
–
E[TAC]
=
$
24
.
6
K
(
new
best,
3
.
4
%
improvement
)
–
E[CT]
for
P
1
=
1
.
94
hours
(
12
%
improvement
)
Solution of the Redesigned Process
P.Y. Yin & Fred Glover
34
•
Simulation Optimization with OptQuest is able to
find high

quality solutions in reasonable time and
re

optimizes the redesigned model
•
These applications are
only a fraction
of the ways
that
metaheuristics and simulation
are used in
optimization involving
non

linearity and
uncertainty
•
Over 60,000 user licenses
of the system have
been sold (each licensed user might have multiple
kinds of problems)
Simulation Optimization with Metaheuristic
P.Y. Yin & Fred Glover
35
•
The genetic metaphors served by most EAs
today have caused
valuable alternative strategies
from human problem solving to be overlooked
•
The human brain and its higher level processes
already exist. The adaptive memory strategies
directly take advantage of human intelligence
that creates another dimension to adapt to
problem solving environment
Conclusions
P.Y. Yin & Fred Glover
36
•
The successes by integrating scatter search and
its path relinking extensions with tabu search
disclose potential advantages for evolutionary
algorithms that incorporate adaptive memory
•
It’s time for a new generation of evolutionary
algorithms
–
Cyber

Evolutionary Algorithms
–
that are evolutionary processes based on
neighborhood structures, adaptive memory, and
responsive strategies
Conclusions
Comments 0
Log in to post a comment