Artificial Intelligence, Logic Programming and Argumentation Group

ADVANCED PROBLEM SOLVING

AND SEARCH
Lecture 4 Tabu Search

Sarah Gaggl

Dresden
Agenda
1 Introduction
2 Uninformed Search versus Informed Search (Best First Search, A*
Search, Heuristics)
3 Local Search, Stochastic Hill Climbing, Simulated Annealing
4 Tabu Search
5 Evolutionary Algorithms/ Genetic Algorithms
6 Answer-set Programming (ASP)
7 Constraint Satisfaction (CSP)
8 Structural Decomposition Techniques (Tree/Hypertree Decompositions)

TU Dresden APSS slide 2 of 30

Tabu Search

Main Idea
• A memory forces the search to explore new areas of the search space
• Memorize solutions that have been examined recently. They become
tabu points in next steps
• Tabu search is deterministic

TU Dresden APSS slide 3 of 30

Tabu Search and SAT
• SAT problem with n = 8 variables
• Initial (random) assignment x = (0, 1, 1, 1, 0, 0, 0, 1)
• Evaluation function: weighted sum of number of satisfied clauses.
Weights depend on the number of variables in the clause
• Maximize evaluation function (i.e. we’re trying to satisfy all clauses)
• Random assignment provides eval(x) = 27
• Neighborhood of x consists of 8 solutions. Evaluate them and select best
• At this stage, it is the same as hill-climbing
• Suppose flipping 3rd variable generates best evaluation (eval(x0 ) = 31)
• Memory keeps track of actions

TU Dresden APSS slide 4 of 30

Recency-based Memory
• Index of flipped variable + time when it was flipped
• Differentiate between older and more recent flips
• SAT: time stamp for each position of solution vector M (initialized to 0)
• Value of time stamp provides information on recency of flip at position

Memory Vector
M(i) = j (when j 6= 0)
j is most recent iteration when i-th bit was flipped

TU Dresden APSS slide 5 of 30

Recency-based Memory
• Index of flipped variable + time when it was flipped
• Differentiate between older and more recent flips
• SAT: time stamp for each position of solution vector M (initialized to 0)
• Value of time stamp provides information on recency of flip at position

Memory Vector
M(i) = j (when j 6= 0)
j is most recent iteration when i-th bit was flipped

Assume information is stored for at most 5 iterations.

Alternative Interpretation
M(i) = j (when j 6= 0)
i-th bit was flipped 5 − j iterations ago

TU Dresden APSS slide 6 of 30

Recency-based Memory
• Index of flipped variable + time when it was flipped
• Differentiate between older and more recent flips
• SAT: time stamp for each position of solution vector M (initialized to 0)
• Value of time stamp provides information on recency of flip at position

Memory Vector
M(i) = j (when j 6= 0)
j is most recent iteration when i-th bit was flipped

Assume information is stored for at most 5 iterations.

Alternative Interpretation
M(i) = j (when j 6= 0)
i-th bit was flipped 5 − j iterations ago

Example
0 0 5 0 0 0 0 0

Memory after one iteration. 3rd bit is tabu for next 5 iterations.

TU Dresden APSS slide 7 of 30

Different Interpretations

1st Variant
• Stores iteration number of most recent flip
• Requires a current iteration counter t which is compared with memory
values
• If t − M(i) > 5 forget
• Only requires updating a single entry, and increase the counter
• Used in most implementations

TU Dresden APSS slide 8 of 30

Different Interpretations

1st Variant
• Stores iteration number of most recent flip
• Requires a current iteration counter t which is compared with memory
values
• If t − M(i) > 5 forget
• Only requires updating a single entry, and increase the counter
• Used in most implementations

2nd Variant
• Values are interpreted as number of iterations for which a position is not
available
• All nonzero entries are decreased by one at every iteration

TU Dresden APSS slide 9 of 30

Example ctd.
• Initial assignment x = (0, 1, 1, 1, 0, 0, 0, 1)
• After 4 additional iterations M :

3 0 1 5 0 4 2 0

• Most recent flip M(4) = 5

• Current solution: x = (1, 1, 0, 0, 0, 1, 1, 1) with eval(x) = 33

TU Dresden APSS slide 10 of 30

Example ctd.
• Initial assignment x = (0, 1, 1, 1, 0, 0, 0, 1)
• After 4 additional iterations M :

3 0 1 5 0 4 2 0

• Most recent flip M(4) = 5

• Current solution: x = (1, 1, 0, 0, 0, 1, 1, 1) with eval(x) = 33

Neighborhood of x

x1 = (0, 1, 0, 0, 0, 1, 1, 1) x5 = (1, 1, 0, 0, 1, 1, 1, 1)
x2 = (1, 0, 0, 0, 0, 1, 1, 1) x6 = (1, 1, 0, 0, 0, 0, 1, 1)
x3 = (1, 1, 1, 0, 0, 1, 1, 1) x7 = (1, 1, 0, 0, 0, 1, 0, 1)
x4 = (1, 1, 0, 1, 0, 1, 1, 1) x8 = (1, 1, 0, 0, 0, 1, 1, 0)

TU Dresden APSS slide 11 of 30

Example ctd.
• Initial assignment x = (0, 1, 1, 1, 0, 0, 0, 1)
• After 4 additional iterations M :

3 0 1 5 0 4 2 0

• Most recent flip M(4) = 5

• Current solution: x = (1, 1, 0, 0, 0, 1, 1, 1) with eval(x) = 33

Neighborhood of x

x1 = (0, 1, 0, 0, 0, 1, 1, 1) x5 = (1, 1, 0, 0, 1, 1, 1, 1)
x2 = (1, 0, 0, 0, 0, 1, 1, 1) x6 = (1, 1, 0, 0, 0, 0, 1, 1)
x3 = (1, 1, 1, 0, 0, 1, 1, 1) x7 = (1, 1, 0, 0, 0, 1, 0, 1)
x4 = (1, 1, 0, 1, 0, 1, 1, 1) x8 = (1, 1, 0, 0, 0, 1, 1, 0)

TABU, best evaluation eval(x5 ) = 32, decrease!

TU Dresden APSS slide 12 of 30

Example ctd.
• Current solution: x = (1, 1, 0, 0, 0, 1, 1, 1) with eval(x) = 33
• New solution: x5 = (1, 1, 0, 0, 1, 1, 1, 1) with eval(x5 ) = 32

3 0 1 5 0 4 2 0

changes to:

2 0 0 4 5 3 1 0

TU Dresden APSS slide 13 of 30

Example ctd.
• Current solution: x = (1, 1, 0, 0, 0, 1, 1, 1) with eval(x) = 33
• New solution: x5 = (1, 1, 0, 0, 1, 1, 1, 1) with eval(x5 ) = 32

3 0 1 5 0 4 2 0

changes to:

2 0 0 4 5 3 1 0

Policy might be too restrictive

• What if tabu neighbor x6 provides excellent evaluation score?
• Make search more flexible: override tabu classification if solution is
outstanding
=⇒ aspiration criterion

TU Dresden APSS slide 14 of 30

Long-term Memory

Question
1 What is stored in long-term memory (think of SAT as an example)?
2 How can we escape local optima with help of a long-term memory?

TU Dresden APSS slide 15 of 30

Frequency-based Memory
• Operates over a longer horizon
• SAT: vector H serves as long-term memory.
– Initialized to 0, at any stage of the search

H(i) = j

interpreted as: during last h (horizon) iterations, the i-th bit was
flipped j times
– Usually horizon is large
– After 100 iterations with h = 50, long-term memory H might have
the following values
5 7 11 3 9 8 1 6
– Shows distribution of moves throughout the last 50 iterations

Diversity of Search
Frequency-based memory provides information about which flips have been
under-represented or not represented.
=⇒ we can diversify the search by exploring these possibilities

TU Dresden APSS slide 16 of 30

Use of Long-term Memory

Special Circumstances
• Situations where all non-tabu moves lead to worse solution
• To make a meaningful decision about which direction to explore next
• Typically: most frequent moves are less attractive
• Value of evaluation score is decreased by some penalty measure that
depends on frequency, final score implies the winner

TU Dresden APSS slide 17 of 30

Example SAT
• Assume value of current solution is eval(x) = 35
• Non-tabu flips 2, 3 and 7 have values 30, 33, 31
• None of tabu moves provides value greater than 37 (highest value so far)
=⇒ we can’t apply aspiration criterion

TU Dresden APSS slide 18 of 30

Example SAT
• Assume value of current solution is eval(x) = 35
• Non-tabu flips 2, 3 and 7 have values 30, 33, 31
• None of tabu moves provides value greater than 37 (highest value so far)
=⇒ we can’t apply aspiration criterion
• Frequency based-memory and evaluation function for new solution x0 is

eval(x0 ) − penalty(x0 )

• penalty(x0 ) = 0.7 × H(i), where 0.7 coefficient, H(i) value from long-term
memory H :

7 for solution created by flipping 2nd bit

11 for solution created by flipping 3nd bit
1 for solution created by flipping 7nd bit

TU Dresden APSS slide 19 of 30

Example SAT
• Assume value of current solution is eval(x) = 35
• Non-tabu flips 2, 3 and 7 have values 30, 33, 31
• None of tabu moves provides value greater than 37 (highest value so far)
=⇒ we can’t apply aspiration criterion
• Frequency based-memory and evaluation function for new solution x0 is

eval(x0 ) − penalty(x0 )
• penalty(x0 ) = 0.7 × H(i), where 0.7 coefficient, H(i) value from long-term
memory H :

7 for solution created by flipping 2nd bit

11 for solution created by flipping 3nd bit
1 for solution created by flipping 7nd bit

• New scores are:

30 − 0.7 × 7 = 25.1 2nd bit

33 − 0.7 × 11 = 25.3 3nd bit
31 − 0.7 × 1 = 30.3 7th bit

TU Dresden APSS slide 20 of 30

Example SAT
• Frequency based-memory and evaluation function for new solution x0 is

eval(x0 ) − penalty(x0 )
• penalty(x0 ) = 0.7 × H(i), where 0.7 coefficient, H(i) value from long-term
memory H :
7 for solution created by flipping 2nd bit
11 for solution created by flipping 3nd bit
1 for solution created by flipping 7nd bit

• New scores are:

30 − 0.7 × 7 = 25.1 2nd bit

33 − 0.7 × 11 = 25.3 3nd bit
31 − 0.7 × 1 = 30.3 7th bit

Diversify Search
Including frequency values in a penalty measure for evaluating solutions.

TU Dresden APSS slide 21 of 30

Further Options to Diversify Search
We migth add additional rules:
• Aspiration by default: select the oldest of all considered
• Aspiration by search direction: memorize whether or not the performed
moves generated any improvement
• Aspiration by influence: measures the degree of change of the new
solution
a) in terms of the distance between old and new solution
b) change in solution’s feasibility, if we deal with a constraint problem
– Intuition: particular move has a larger influence if a larger step was
made from old to new solution

TU Dresden APSS slide 22 of 30

Tabu Search and the TSP
• Move: swap two cities in a particular solution
• Current solution: (2, 4, 7, 5, 1, 8, 3, 6)
• 28 neighbors 8 = 7·8 = 28


2 2
• Recency-based memory: swap of cities i and j in i-th row and j-th column
(for i < j)
• Maintain number of remaining iterations for which swap stays on tabu list
• Frequency-based memory: same structure; indicate totals of all swaps
within horizon h = 50

TU Dresden APSS slide 23 of 30

Tabu Search and the TSP
• Move: swap two cities in a particular solution
• Current solution: (2, 4, 7, 5, 1, 8, 3, 6)
• 28 neighbors 82 = 7·8


2
= 28
• Recency-based memory: swap of cities i and j in i-th row and j-th column
(for i < j)
• Maintain number of remaining iterations for which swap stays on tabu list
• Frequency-based memory: same structure; indicate totals of all swaps
within horizon h = 50
2 3 4 5 6 7 8
1
2
3
4
5
6
7

TU Dresden APSS slide 24 of 30

Tabu Search and the TSP ctd.
• Assume both memories initialized to zero and 500 iterations have been
completed
• Current solution: (7, 3, 5, 6, 1, 2, 4, 8) with length: 173, best solution so far
171

2 3 4 5 6 7 8 2 3 4 5 6 7 8
0 0 1 0 0 0 0 1 0 2 3 3 0 1 1 1
0 0 0 5 0 0 2 2 1 3 1 1 0 2
0 0 0 4 0 3 2 3 3 4 0 3
3 0 0 0 4 1 1 2 1 4
0 0 2 5 4 2 1 5
0 0 6 3 1 6
0 7 6 7

left: recency-based memory; right: frequency-based memory

TU Dresden APSS slide 25 of 30

Particular Implementation of Tabu Search
for TSP

Algorithm Tabu Search for TSP [Knox, J. (1994)]

tries ← 0
repeat
generate a tour
count ← 0
repeat
identify a set T of 2-interchange moves
select the best admissible move from T
make appropriate 2-interchange
update taub list and other variables
if the new tour is the best-so-far for a given ’tries’ then
update local best tour information
else
count ← count + 1
end if
until count = ITER
tries ← tries + 1
if the current tour is the best-so-far (for all ’tries’) then
update global best tour information
end if
until tries = MAX-TRIES

TU Dresden APSS slide 26 of 30

Particular Implementation of Tabu Search
for TSP ctd.
• A tour is tabu if both added edges of interchange are on tabu list
• Tabu list update: placing deleted edges on list
• Tabu list is of fixed size
• Whenever it is full, the oldest element in list is replaced by new deleted
edge
• Initially, list is empty and all elements of aspiration list are set to large
values
• Note: Algorithm examines all neighbors, i.e. all 2-interchange tours

TU Dresden APSS slide 27 of 30

Particular Implementation of Tabu Search
for TSP ctd.
• A tour is tabu if both added edges of interchange are on tabu list
• Tabu list update: placing deleted edges on list
• Tabu list is of fixed size
• Whenever it is full, the oldest element in list is replaced by new deleted
edge
• Initially, list is empty and all elements of aspiration list are set to large
values
• Note: Algorithm examines all neighbors, i.e. all 2-interchange tours
• Best results were achieved when
– Length of tabu list was 3n (n number of cities of problem)
– Candidate tour could override tabu status if both edges passed
aspiration test. Compared length of tour with aspiration values for
both added edges. If length of tour was better than both aspiration
values, the test was passed.
– Values present on aspiration list were tour costs prior to the
interchange
– MAX-TRIES and ITER (# of interchanges) depend on size of the
problem. For 100 cities or less, MAX-TRIES was 4, and ITER was
set to 0.0003 · n4 .
TU Dresden APSS slide 28 of 30
Summary
• Simulated annealing and tabu search are both design to escape local
optima
• Tabu search makes uphill moves only when it is stuck in local optima
• Simulated annealing can make uphill moves at any time
• Simulated annealing is stochastic, tabu search is deterministic
• Compared to classic algorithms, both work on complete solutions. One
can halt them at any iteration and obtain a possible solution
• Both have many parameters to worry about

TU Dresden APSS slide 29 of 30

References
Zbigniew Michalewicz and David B. Fogel.
How to Solve It: Modern Heuristics, volume 2. Springer, 2004.
Knox, J.
Tabu Search Performance on the Symmetric Traveling Salesman
Problem, Computer Operations Research, Vol.21, No.8, pp.867–876,
1994.

TU Dresden APSS slide 30 of 30