University of Zagreb, Croatia

Lecture 7
Heuristic Optimization Methods:
Tabu Search

Slides prepared by Nina Skorin-Kapov

Academic year 2020/2021

Graduate Studies Programme
 Outline

 Tabu Search concept

 TSdimensions
 Memory types

 Main design issues

 Tabu list
 Long-term memory
 Aspiration Criterion

 Pseudocode
 TS design elements

Zavod za telekomunikacije 2
 Optimization methods
Exact Approximate
A*
methods methods
Simplex- Constraint
based for LP programming Approximation
algorithms
Heuristic
Branch and Bound, Dynamic
Cut , Price (B&B, programming algorithms
B&C, B&P)

Improvement (meta)heuristics
Constructive heuristics
• Local search
• Greedy algorithms
•Nature inspired
• Tabu search, Genetic
Hybrid methods
algorithms, Simmulated
• GRASP Annealing, Ant colony
•Problem specific
heuristics

Zavod za telekomunikacije 3
 Tabu search

 Iterative, single-solution based metaheuristic

 Developed in the 1970’s by Fred Glover
 Became very popular in the 1990’s
 Introduces the idea of memory to the search
 MAIN IDEA: use knowledge of previous experiences to
make better decisions; guides a local search to explore
beyond local optimality
 Iteratively searches neighborhoods in the solution
space as in local search, but an adaptive memory
guides the search

Zavod za telekomunikacije 4
 Tabu Search: Escaping local optima

initial solution

Tabu search:
Local Search Allows non
solution: the first improving
local optimum solutions

Zavod za telekomunikacije 5
 Tabu List: Avoid cycling

initial solution Tabu list:

previously
visited solutions The search
are tabu moves forward!

After leaving the

local optimum,
the new best
local optimum
? neighboring
Tabu search: solution is the
Allow non local optimum
improving again?
solutions
Zavod za telekomunikacije 6
 Tabu search run
 The search moves forward for a desired number of
iterations; best seen (incumbent) solution is always kept

initial solution

Optimal
solution, but the Final solution: Not
search did not best seen necessarily
find it (incumbant) optimal!
Heuristic!
Zavod za telekomunikacije 7
 Memory types

Explicit • Records the complete solution

• Mostly used for recording elite
memory solutions

• Records part of the solution, an

attribute, which changes in moving
Attributive from one solution to another
memory • Records the transformation (move)
• Mostly used for guiding the search

Zavod za telekomunikacije 8
 Example: explicit memory

max f ( x) = x3 − 60 x 2 + 900 x + 50
x = [0,31], integer
 Representation: x is shown in Explicit
Explicit
binary form with 5 bits memory
memory
s 1 0 0 1 1
ITERATION 1: f(s)
0 0 0 1 1 2237
Memorize full
1 1 0 1 1 293
N(s) previous solution
1 0 1 1 1 1177
1 0 0 0 1 2923 1 0 0 1 1
1 0 0 1 0 2642
Zavod za telekomunikacije 9
 Example: attributive memory

max f ( x) = x3 − 60 x 2 + 900 x + 50
x = [0,31], integer
 Representation: x is shown in Attributive
binary form with 5 bits memory
s 1 0 0 1 1
ITERATION 1: f(s)
0 0 0 1 1 2237 Memorize move
1 1 0 1 1 293 from previous to
N(s)
1 0 1 1 1 1177 new solution
1 0 0 0 1 2923
1 0 0 1 0 2642 Flip 4th bit

Zavod za telekomunikacije 10
 Tabu List (1/2)

 A memory structure simulating

short term memory
 Records a certain number of
recently (last) visited solutions or
Tabu List
attributes/moves
 These solutions or
atributes/moves are forbidden
(tabu) for as long as they remain
in the list
 E.g. the tabu list is of size 3 updated
cyclically after each iteration →
forbidden for 3 iterations
Zavod za telekomunikacije 11
 Tabu List (2/2)

 Used to avoid getting stuck in

local optima and cycling between
neighboring solutions
 Proceeds as local search but Tabu List
accepts non-improving solutions
 When the search gets to a local
optimum, the search proceeds with a
non-improving solution not in the tabu list
 No backtracking
 A certain number of previously visited
solutions (in the tabu list) are tabu
which avoids cycling between
neighboring solutions

Zavod za telekomunikacije 12
 Tabu list: structure and length
 The tabu list is a problem-specific design issue
 The structure depends on whether it uses
explicit/attribute memory
 Usually updated cyclically after each iteration
 Tabu tenure: for how many iterations do the
solutions/attributes in the list remain forbidden
(tabu)
 Ifthe structure is such that explicit solutions are stored,
the tenure is the length of tabu list
 Problem-specific; often determined experimentally
(parameter tuning)
 If too big→the search can become too restrictive and inefficient
 If too small→the search may get stuck in a local area
Zavod za telekomunikacije 13
 Example
 Objective and fitness function:
3 2
max f ( x) = x − 60 x + 900 x + 50
x = [0,31], integer
 Representation: x in binary form with 5 bits
 Initial solution:

1 0 0 1 1 f(s)=2349

 Neighborhood: “1-flip”
 Termination criterion: 6 iterations
Zavod za telekomunikacije 14
 Example

 Use only short -term attributive memory

 Tabu list: records position of last changed bit
 Tabu tenure: 3, i.e. tabu moves are tabu for 3
iterations
 Structure of tabu list: a vector of 5 integers with
values in {0,1,2,3} each corresponding to one bit in
the solution representation and each indicating for
how many more iterations is changing that bit tabu

1st bit 5th bit

Tabu list 0 0 0 0 0 Initially tabu list empty– zero vector

Zavod za telekomunikacije 15
 Example

Current
1 0 0 1 1 f(s)=2349
solution s Tabu list

Incumbent 0 0 0 0 0
1 0 0 1 1 f(sbest)=2349
solution sbest

ITERATION 1: Neighborhood “1 flip”

f(s)
0 0 0 1 1 2237 Best neighbor:
1 1 0 1 1 293 f(s’)=2923
N(s) New current
1 0 1 1 1 1177
1 0 0 0 1 2923 solution
1 0 0 1 0 2642
Zavod za telekomunikacije 16
 Example
ITERATION 1: f(s)
0 0 0 1 1 2237 Best neighbor: f(s’)=2923
1 1 0 1 1 293
N(s) 1 0 1 1 1 1177
1 0 0 0 1 2923 s 1 0 0 0 1 f(s)=2923
1 0 0 1 0 2642 New current solution
Obtained from old
current solution by Update Tabu list 0 0 0 3 0
changing the 4th bit tabu list
Is s better than sbest, i.e. is
f(s)>f(sbest)? Update
incumbent sbest 1 0 0 0 1
f(s)=2923>f(sbest)=2349
YES! solution f(sbest)=2923

Zavod za telekomunikacije 17
 Example
After ITERATION 1:
s 1 0 0 0 1 f(s)=2923

Tabu list 0 0 0 3 0
Generate
neighborhood sbest 1 0 0 0 1 f(sbest)=2923

ITERATION 2: f(s)
0 0 0 0 1 891
1 1 0 0 1 675 Best neighbor not
N(s) tabu: f(s’)=3186
1 0 1 0 1 1751
tabu - New current
1 0 0 0 0 3186 solution

Zavod za telekomunikacije 18
 Example
ITERATION 2: f(s)
Best neighbor: f(s’)=2923
0 0 0 0 1 891
1 1 0 0 1 675
N(s)
1 0 1 0 1 1751
s 1 0 0 0 0 f(s)=3186
tabu -
1 0 0 0 0 3186 New current solution
Obtained from old
current solution by Update Tabu list 0 0 0 2 3
changing the 5th bit tabu list

Is s better than sbest, i.e. is

Update
f(s)>f(sbest)? sbest 1 0 0 0 0
f(s)=3186>f(sbest)=2923 incumbent
YES! solution f(sbest)=3186

Zavod za telekomunikacije 19
 Example
After ITERATION 2:
s 1 0 0 0 0 f(s)=3186

Tabu list 0 0 0 2 3
Generate
neighborhood sbest 1 0 0 0 0 f(sbest)= 3186

ITERATION 3: f(s)
0 0 0 0 0 50
1 1 0 0 0 914 Best neighbor not
N(s) tabu: f(s’)=2050
1 0 1 0 0 2050
tabu - New current
tabu - solution

Zavod za telekomunikacije 20
 Example

ITERATION 3: f(s)
Best neighbor: f(s’)=2050
0 0 0 0 0 50
N(s) 1 1 0 0 0 914
1 0 1 0 0 2050 s 1 0 1 0 0 f(s)=2050
tabu -
tabu - New current solution
Obtained from old
Update Tabu list
current solution by 0 0 3 1 2
tabu list
changing the 3rd bit
Is s better than sbest, i.e. is
f(s)>f(sbest)? Don’t update sbest 1 0 0 0 0
f(s)=2050>f(sbest)=3186 incumbent
NO! solution f(sbest)=3186

Zavod za telekomunikacije 21
 Example

After ITERATION 3:
s 1 0 1 0 0 f(s)=2050

Tabu list 0 0 3 1 2
Generate
neighborhood sbest 1 0 0 0 0 f(sbest)=3186

ITERATION 4: f(s)
0 0 1 0 0 2754 Best neighbor not
N(s) 1 1 1 0 0 162 tabu: f(s’)=2754
tabu - New current
tabu - solution
tabu -
Zavod za telekomunikacije 22
 Example
ITERATION 4: f(s)
0 0 1 0 0 2754 Best neighbor: f(s’)=2754
N(s) 1 1 1 0 0 162
tabu -
tabu - s 0 0 1 0 0 f(s)=2754
tabu - New current solution
Obtained from old
Update Tabu list
current solution by 3 0 2 0 1
tabu list
changing the 1st bit
Is s better than sbest, i.e. is
f(s)>f(sbest)? Don’t update sbest 1 0 0 0 0
f(s)=2754>f(sbest)=3186 incumbent
NO! solution f(sbest)=3186

Zavod za telekomunikacije 23
 Example

After ITERATION 4:
s 0 0 1 0 0 f(s)=2754

Tabu list 3 0 2 0 1
Generate
neighborhood sbest 1 0 0 0 0 f(sbest)=3186

ITERATION 5: f(s)
tabu -
0 1 1 0 0 3938 Best neighbor not
N(s) tabu: f(s’)=3938
tabu -
0 0 1 1 0 3506 New current
tabu - solution

Zavod za telekomunikacije 24
 Example
ITERATION 5: f(s)
Best neighbor: f(s’)=3938
tabu -
0 1 1 0 0 3938
N(s)
tabu -
s 0 1 1 0 0 f(s)=3938
0 0 1 1 0 3506
tabu - New current solution
Obtained from old
Update Tabu list
current solution by 2 3 1 0 0
tabu list
changing the 2nd bit
Is s better than sbest, i.e. is
f(s)>f(sbest)? Update sbest 0 1 1 0 0
f(s)=3938>f(sbest)=3186 incumbent
YES! solution f(sbest)=3938

Zavod za telekomunikacije 25
 Example

After ITERATION 5:
s 0 1 1 0 0 f(s)=3938

Tabu list 2 3 1 0 0
Generate
neighborhood sbest 0 1 1 0 0 f(sbest)=3938

ITERATION 6: f(s)
tabu -
tabu - Best neighbor not
N(s) tabu: f(s’)=3807
tabu -
0 1 1 1 0 3634 New current
0 1 1 0 1 3807 solution

Zavod za telekomunikacije 26
 Example
ITERATION 6: f(s)
Best neighbor: f(s’)=3807
tabu -
tabu -
N(s)
tabu -
s 0 1 1 0 1 f(s)=3807
0 1 1 1 0 3634
0 1 1 0 1 3807 New current solution
Obtained from old
Update Tabu list
current solution by 1 2 0 0 3
tabu list
changing the 5th bit
Is s better than sbest, i.e. is
f(s)>f(sbest)? Don’t update sbest 0 1 1 0 0
f(s)=3807>f(sbest)=3938 incumbent
NO! solution f(sbest)=3938

Zavod za telekomunikacije 27
 Example
After ITERATION 6:
s 0 1 1 0 1 f(s)=3807

Tabu list 1 2 0 0 3

sbest 0 1 1 0 0 f(sbest)=3938

Stopping criterion: 6 iterations → STOP! Note: the TS

solution is not
sbest is the best found solution globally optimal
x=01010
sbest 0 1 1 0 0 f(sbest)=3938 f(x)=4050

Zavod za telekomunikacije 28
 Alternative Tabu list structure

 Alternative structure of the same Tabu list

A FIFO buffer of size 3
 Elements in buffer are tabu moves

Tabu list
ITERATION 0 (initally): - - -
After ITERATION 1: 4 - -
After ITERATION 2: 5 4 -
After ITERATION 3: 3 5 4
After ITERATION 4: 1 3 5
After ITERATION 5: 2 1 3
After ITERATION 6: 5 2 1
Zavod za telekomunikacije 29
 Tabu Search dimensions
Memory structures used in TS are designed to aid 4 different
dimensions of the search

Recency Frequency

Quality Influence

Zavod za telekomunikacije 30
 TS dimensions
 Recency
 Short-term memory
 Keeps track of solutions or
attributes/moves which have recently Recency
been visited and prohibits revisiting
them
 Most commonly used memory type
 Tabu list
 If attributes or moves are memorized:
 Selected attributes/changes occurring in
recently visited solutions become tabu,
i.e., all solutions (both already-visited and
not-yet-visited) which have those
attributes are prohibited while in the tabu
list
Zavod za telekomunikacije 31
 TS dimensions

 Frequency
 Long-term memory
 Records how many times we
have returned to the same Frequency
solution or attribute
 Residence measures: the number of
times an attribute/solution is
observed
 Transition measures: the number of
times an attribute/solution changes
from one value to another
 Complements recency-based
memory

Zavod za telekomunikacije 32
 TS dimensions

 Quality
 During the search, it records the
quality of solutions with certain
attributes Quality
 Can be used to identify
attributes/elements that are
common to all (most) ‘good’
solutions → intensification
 Can enable incentive-based
learning: reinforce moves which
lead to good solutions and penalize
those which lead to bad ones
 (often used in aspiration criterion)

Zavod za telekomunikacije 33
 TS dimensions

 Influence
 Records the impact (influence)
choices made during the search
Influence
have on the quality and structure
of the solutions
 Some moves can drastically change
the solutions
 Record how far we have moved
from the previous solution
(distance between solutions)
 (often used in aspiration criterion)

Zavod za telekomunikacije 34
 Main TS-specific design issues

• A short-term memory structure (recency dimension) used to

prevent the search from revisiting recently visited
Tabu list solutions/attributes
• Of finite size, updated after each iteration; memorizing all
previous solutions may confine search and affect efficiency

Basic TS
technique
Long-term • Relates mostly to the frequency dimension
• Used for 2 main different purposes: DIVERSIFICATION
memory and INTENSIFICATION

Advanced TS
techniques
Aspiration • Utilizes the quality and influence dimensions to guide the
search to good solutions
• Can override the tabu list in case of exceptional solutions
criterion (better than incumbent)

Zavod za telekomunikacije 35
 Long-term memory

 Relates mostly to the second TS

dimension: Frequency
 Records the number of times an
Long-term
attribute/solution is observed or memory
changes from one value to another
 Used for 2 main different
purposes:
 DIVERSIFICATION
 Long-term memory
 INTENSIFICATION
 Medium/long-term memory

Zavod za telekomunikacije 36
 Long-term memory purposes

• Guides the search to explore unvisited areas

• If some solutions or neighborhoods are often
searched, i.e. some solutions/attributes have very
Diversification high frequency, continue the search in new area
(to solutions which drastically differ from those
Long-term

often visited)
memory

• If good solutions/attributes (quality, influence)

are often visited (frequency), then intensify the
Intensification search in that area od the solution space
• Give priority to elite solutions or make the
search biased by attributes of elite solutions

Zavod za telekomunikacije 37
 Aspiration criterion

 Relates mostly to TS
dimensions: quality and
influence Aspiration
 Allows overriding the tabu list Criterion
by choosing a tabu move if it
is better than the incumbent
solution
 Can also allow a tabu move
which generated a better
solution in a set of solutions
with a given attribute

Zavod za telekomunikacije 38
 Pseudocode - minimization
Start Tabu Search
Generate initial solution s ;
0
s  s ; s best  s ;
0 0
TL = {}; // tabu list initially empty
Initialize aspiration criterion (AC)
Initialize other memory structures (e.g. long - term) if any;
do
Generate N(s, TL , AC); // neighborho od subject to tabu list and aspiration criterion
Find s ' N ( s, TL , AC ) where f ( s ' )  f ( s ' ' ) , s' ' N(s, TS, AC)
Update s  s '
Update tabu list, s best , aspiration criterion, other memory structures;
If intesifica tion or diversific ation criterion then intesify or diversify search;
while stopping criterion not met
return s best ;
End

Zavod za telekomunikacije 39
 Design components

 Solution representation (e.g. binary vector)

 Initial solution (e.g. random, greedy)
 Objective function
 Fitness function
 Neighborhood definition (‘move’ operator)
 Tabu list structure
 Termination criterion
 Additional components (optional):
 Long/medium-term memory
 Diversification and intesification criteria
 Aspiration criterion
Zavod za telekomunikacije 40