Title: Testing New Ideas for Metaheuristic Search with the MaxCut Problem
1Testing New Ideas for Metaheuristic Search with
the Max-Cut Problem
2References
- Laguna, M., A. Duarte and R. Martí (forthcoming)
Hybridizing the Cross Entropy Method An
Application to the Max-Cut Problem, Computers
and Operations Research. - Martí, R., A. Duarte and M. Laguna (second
revision) Advanced Scatter Search for the
Max-Cut Problem, INFORMS Journal on Computing.
3The Cross Entropy Method
- Due to Rubinstein (1997)
- Adapted for the solution of combinatorial
optimization (Rubinstein 1999 2001) - Recent Annals of OR volume devoted to Cross
Entropy (vol. 134, 2005) - Website http//www.cemethod.org
4Basic Steps
- Generate a random sample from a pre-specified
probability distribution function - Use the sample to modify the parameters of the
probability distribution in order to produce a
better sample in the next iteration
5Example
6CE Pseudo Code
7Search Parameters
8Adding Local Search
9HCE Pseudo Code
10The Max-Cut Problem
V1
j
wij
i
V0
V n 9 E 16
11Applications
- VLSI circuit design
- Telecommunications networks
- Statistical mechanics (spin glasses)
12CE for the Max-Cut Problem
xi 0 if vertex i is assigned to V0 xi 1 if
vertex i is assigned to V1
13Local Search for the Max-Cut
For each variable xi define
Examine all nodes in random order and stop when
no node is reassigned after a complete pass.
14Data Sets for Experimentation
- Rubinstein (2002)
- Six instances with n 200, E 19,900 and 1 ?
w ? 5 - 7th DIMACS Implementation Challenge
- Four instances with 1536 ? E ? 10125
- Helmberg and Rendl (2000)
- Twenty four instances with 800 ? n ? 3000 and
density between 0.17 to 6.12 - Festa et al. (2002)
- Twenty instances, ten with n 1000 and 0.6
density and ten with n 2744 and 0.22 density
15Parameter Fine-tuning
- We could not reproduce the results reported by
Rubinstein (2002) - Calibra (Adenso-Diaz and Laguna 2007) was used to
fine tune both CE and HCE
Adenso-Díaz, B. and M. Laguna (2006) Fine-tuning
of Algorithms Using Partial Experimental Designs
and Local Search, Operations Research, vol. 54,
no. 1, pp. 99-114.
16Results for Set 1
Single run with same random seed
17Results for Set 2
18Results for Sets 3 and 4
19SS Implementation
- Solution of the maximum diversity problem to
increase diversity in the reference set - Dynamic adjustment of a key parameter within the
search - Adaptive selection of a combination method
20Relevant Max-Cut Literature
21Scatter Search Framework
22Diversification Generation
- C1 Method by Festa et al (2002)
- GRASP construction that starts with one node in
V0 and one in V1 and uses sigma values to
iteratively assign the rest - C2 Method
- GRASP construction that starts with all nodes
assigned to V0 and iteratively moves nodes to V1
to maximize the value of the cut - C3 Method
- Similar to C2 but uses frequency memory instead
of randomization
23Improvement Method
- LS1 by Festa
- Uses sigma values to identify improving moves and
uses a best move strategy - LS2 by Festa
- Uses the notion of ejection chains and a first
improving strategy
24Combination Method
- CB1
- Uses weighted scores to bias a probabilistic
assignment of the nodes - CB2
- Finds common elements and then uses sigma values
and randomness to produce to solutions - CB3
- Based on path relinking proposed by Festa et al
(2002)
25New Elements
- Solve the maximum diversity problem to construct
first reference set - Include ejection chain parameter k in the search
process - Combination methods are probabilistically
selected according to their relative merit
26Experiments with Construction Methods
27Experiments with Improvement Methods
C3LS2 variants are statistically equivalent
28Experiments with Combination Methods
29Contribution to Best Solution
30SS vs. CirCut
31Observations
- Basic CE cant compete with specialized methods
- HCE is a general framework that uses a
context-specific local search - More experimentation is necessary to determine
the merit of the suggested new elements for SS