Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent...

30
Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne

Transcript of Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent...

Page 1: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP)

Tim Daniëlse en Vincent Delvigne

Page 2: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

QAP

• facilities (items)• locations• distancematrix A (between locations)• flowmatrix B (between facilities)• Assign facilities to locations, minimizing

transportation needed.

Page 3: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

QAP (2)

• is the distance between locations and • is the flow between facilties and • is the set of all permutations of • gives the location of item in solution

Page 4: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Complexity

• NP-hard• Largest non-trivial instance solved to

optimality: • Thus use heuristics:– Simulated annealing– Tabu search– Memetic algorithms– Ant algorithms– Scatter search

Page 5: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search

Procedure Iterated Local SearchGenerateInitialSolution LocalSearch()

repeat Perturbation() LocalSearch(’) AcceptanceCriterion()

Until termination condition metEndProcedure

Page 6: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search (2)

• Generate Initial Solution– No known, well performing construction algorithm– Randomized assignment

Page 7: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search (3)

• Local Search– 2-opt– First-improvement pivot

– Don’t look bits

Page 8: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search (4)

• Perturbation– random k-opt– best value of k not known a priori– Adaptive, between and

Page 9: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Iterated Local Search (5)

• Acceptance Criterion– standard criterion: accept only improvements– varies among the algorithm-variants

Page 10: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

QAP instance classes

• QAPLIB benchmark library• 4 instance classes:– randomly generated (class i)– Manhattan distance matrix (class ii)– real-life instances (class iii)– random, resembling real-life (class iv)

Page 11: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

QAP instance classes (2)

• Functions to differentiate amongst classes:- Flow dominance

- Distance dominance

- Sparsity

where is number of “0” entries in A or B

Page 12: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Analysis of search space

• Fitness-Distance correlation analysis

• 5000 LS runs• 1000 ILS runs with iterations.

Page 13: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Analysis of search space (2)

Page 14: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Analysis of search space (3)

Page 15: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Distance-Fitness Correlation

Page 16: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Distance-Fitness Correlation (2)

Page 17: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Stagnation Detection• Empirical run-time

distribution (RTD)• RTD develops below

exponential distribution (stagnation)– perform restart

Page 18: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Variations

• Soft Restarts– Use of history– Random new solution.

• Random Walk (RW)– Accept answer regardless of improvement– Combination with default “Better” might improve

even more

Page 19: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Variations (2)

• Large Step Markov Chains (LSMC)– Accepts worse solution with certain probability.– Similar to simulated annealing – Uses temperature parameter

Page 20: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Variations (3)

• Population Based Extensions• Replace-Worst– Start with solutions each with standard ILS– Every iterations a copy of the best replaces the

worst

Page 21: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Variations (4)

• Evolution strategies (ES)– Population of solutions– Each iteration new solutions are generated– Uses distance to determine membership of

population– Parameter

Page 22: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Comparison

• Robust Tabu search (RoTS)– Best for class and

• Max-Min Ant System (MMAS)– Best for class and

Page 23: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Comparison (2)

Page 24: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Algorithm Comparison (3)

Page 25: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.
Page 26: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Evolutionary Variant

• Variant of the Evolutionary Strategy.

• Optimized Local Search• Different parameter settings

Page 27: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.
Page 28: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Evolutionary Variant (3)

Page 29: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

Conclusions

• Fitness-Distance Correlation analysis• ILS runtime analysis• Acceptance Criteria analysis• ES-MN best performing algorithm.

Page 30: Iterated Local Search (ILS) for the Quadratic Assignment Problem (QAP) Tim Daniëlse en Vincent Delvigne.

References

• Stutzle, Thomas (2006): Iterated Local Search for the Quadratic Assignment Problem. European Journal of Operational Research 174 (3), 1519-1539.