network crossover performance on nk landscapes and deceptive problems
DESCRIPTION
Presentation at GECCO-2010TRANSCRIPT
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover Performance on NKLandscapes and Deceptive Problems
M. Hauschild1 M. Pelikan1
1Missouri Estimation of Distribution Algorithms Laboratory (MEDAL)Department of Mathematics and Computer Science
University of Missouri - St. Louis
Genetic and Evolutionary Computation Conference, 2010
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Motivation
Always looking to solve difficult problems with GAs.In a scalable and robust manner.Must respect linkage between bits.
Most common variation operators do not do this.Uniform, two-point crossover.
One solution is linkage-learning GAs.EDAs respect linkages.Come at the cost of model-building.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Motivation
Often have prior information about a problem.Graph-based problems.EDA models on similar problems.
What is the best way to exploit this information?Bias EDA model building.Sample directly from a network model.Modify the crossover operator itself.
Test this operator against an EDA.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Outline
Network crossoverAlgorithms
GAhBOADeterministic Hill-Climber
Test ProblemsExperiments
Trap-5NK Landscapes
Conclusions
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
Two-parent crossovers start by creating a binary mask.What bits to exchange and what to keep the same.
Uniform crossover sets the bits randomly.How to create a mask to respect linkages?
Start with a matrix G specifying strongest linkages.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
G is often not hard to obtain.Graph problems have this implicitly.MAXSAT and other problems also easy.Trial runs of EDAs.
Only requires strongest connections.Does not require perfect knowledge.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
To build the maskChoose a random bitRandomized breadth-first search to expand maskRepeat until mask is complete
Stop when mask size is n/2.
Bits close in G less likely to be disrupted.
Bits far from each other more likely to be disrupted.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Network Crossover
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Genetic Algorithm
Three crossover operators used.Network crossoverTwo-pointUniform
Probability of crossover, pc = 0.6
Probability of mutation, pm = 1/n
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
hierarchical Bayesian Optimization Algorithm (hBOA)
Pelikan, Goldberg, and Cantú-Paz; 2001Uses Bayesian network with local structures to modelsolutions
Acyclic directed GraphString positions are the nodesEdges represent conditional dependenciesWhere there is no edge, implicit independence
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Deterministic Hill-Climber
Deterministic hill climber (DHC) used for all runsPerforms single-bit changes that lead to maximumperformanceStops when no single-bit change leads to improvement
Originally considered not using DHCDramatically improved performance
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Trap-5
Partition binary string into disjoint groups of 5 bits
trap5(ones) =
{
5 if ones = 54 − ones otherwise
, (1)
Total fitness is sum of single traps
Global Optimum: String 1111...1
Local Optimum: 00000 in any partition
G has all bits in the same partition connected
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
NK Landscapes
Popular test function developed by Kaufmann (1989).
Gives a model of a tunable rugged landscape.An NK fitness landscape is defined by
Number of bits, n.Number of neighbors per bit, k .Set of k neighbors
∏
(Xi ) for i-th bit, Xi .Subfunction fi defining contribution of Xi and
∏
(Xi ).
The objective function fnk to maximize is defined as
fnk (X0, . . . , Xn−1) =
n−1∑
i=0
fi(Xi ,∏
(Xi)) (2)
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
NK landscapes
Nearest neighbor NK landscapes.Bits are arranged in a circle.Neighbors of each bit restricted to the following k bits.Parameter step ∈ {1, 2, . . . , k + 1} used to control overlap.
For step = 1, maximum overlap.For step = k + 1, fully separable.
Bit positions shuffled randomly to increase difficulty.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
NK landscapes
Unrestricted NK landscapes.NP-complete for k > 1Branch and bound algorithm used to find optima.
Nearest neighbor NK landscapes.Polynomial solvability.Dynamic programming used to find optima.
G connects all neighboring bits.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Experimental Setup
Trap-5Problem sizes from n = 100 to n = 300.Bisection used, 10 out of 10 independent runs.10 independent bisection runs performed.Some experiments cut short at extreme problem sizes.
Unrestricted NK landscapesProblem sizes of n ∈ {20, 22, . . . , 38}.k = 51000 random problem instances for each setting.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Experimental Setup
Nearest neighbor NK landscapesProblem sizes of n ∈ {30, 60, . . . , 210}.Two step sizes considered, step ∈ {1, 5}.k = 51000 instances for each combination of n, k , step.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Experimental Setup
Two replacement techniques considered.Restricted Tournament Replacement(RTR)
Niching, replaces similar solutions.Window size set to w = min{n, N/5}.
ElitismKeeps a portion of the best individuals each generation.50% of the most fit individuals kept.
Examined three measuresEvaluationsLocal search stepsExecution time
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Trap-5
Evaluations, RTR
100 150 200 250 300
105
Problem Size
Eva
luat
ions
netxuniformhboa2−point
DHC flips, RTR
100 150 200 250 300
105
1010
Problem Size
Num
ber
of fl
ips
netxuniformhboa2−point
Execution Time, RTR
100 150 200 250 300
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2−point
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Trap-5
Evaluations, elitism
100 150 200 250 300
105
Problem Size
Eva
luat
ions
netxuniformhboa
DHC flips, elitism
100 150 200 250 300
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa
Execution Time, elitism
100 150 200 250 300
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK, step = 5
Evaluations, RTR
30 60 90 120150 210
104
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, RTR
30 60 90 120150 210
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, RTR
30 60 90 120150 210
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK, step = 5
Evaluations, RTR
30 60 90 120150 210
104
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, RTR
30 60 90 120150 210
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, RTR
30 60 90 120150 210
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK, step = 1
Evaluations, RTR
30 60 90 120150 210
104
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, RTR
30 60 90 120150 210
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, RTR
30 60 90 120150 210
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK, step = 5
Evaluations, elitism
30 60 90 120150 210
105
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, elitism
30 60 90 120150 210
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, elitism
30 60 90 120150 210
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK, step = 1
Evaluations, elitism
30 60 90 120150 210
105
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, elitism
30 60 90 120150 210
105
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, elitism
30 60 90 120150 210
100
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Nearest neighbor NK by difficulty
n = 120, step = 5 n = 120, step = 1
0 0.5 10.2
0.4
0.6
0.8
1
Percent easiest netx instances
Num
ber
of fl
ips/
mea
n
netxuniformhboa
0 0.5 10.2
0.4
0.6
0.8
1
Percent easiest netx instances
Num
ber
of fl
ips/
mea
n
netxuniformhboa
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Unrestricted NK landscapes
Evaluations, RTR
20 22 26 30 34 3810
2
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, RTR
20 22 26 30 34 38
103
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, RTR
20 22 26 30 34 38
10−2
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Unrestricted NK landscapes
Evaluations, elitism
20 22 26 30 34 3810
2
Problem Size
Eva
luat
ions
netxuniformhboa2p
DHC flips, elitism
20 22 26 30 34 38
103
Problem Size
Num
ber
of fl
ips
netxuniformhboa2p
Execution Time, elitism
20 22 26 30 34 38
10−2
Problem Size
Exe
cutio
n T
ime
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Unrestricted NK by difficulty
n = 38, RTR n = 38, elitism
0 0.5 10.2
0.4
0.6
0.8
1
Percent easiest netx instances
Num
ber
of fl
ips/
mea
n
netxuniformhboa2p
0 0.5 10.2
0.4
0.6
0.8
1
Percent easiest netx instances
Num
ber
of fl
ips/
mea
n
netxuniformhboa2p
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Conclusions
Compared GA with network crossover againstGA with uniform and two-point crossover.hBOA, a state of the art EDA.
On nearest neighbor NK landscapes and trap5Network crossover had the best execution time through allsettings.Niching with RTR outperformed elitism.hBOA had the least variance in instance difficulty.
On unrestricted NK landscapesResults less clear.hBOA had the best scalability.RTR and elitism results were mixed.
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Conclusions
Future WorkTest on more diverse problems.Use trial runs of an EDA to learn the crossover network.Test other network based crossovers.Test against a version of hBOA that takes into accountproblem structure
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems
Motivation Outline Network Crossover Algorithms Test Problems Experiments Conclusions
Any Questions?
M. Hauschild and M. Pelikan University of Missouri - St. Louis
Network Crossover Performance on NK Landscapes and Deceptive Problems