a schema-based evolutionary alg’m. for black-box optimization
DESCRIPTION
A Schema-Based Evolutionary Alg’m. for Black-Box Optimization. David A. Cape CS 448, Spring 2008 Missouri S & T. Motivation. Arbitrary Additively Decomposable Functions Example: multivariate polynomial (sum of two 4-bit D-Traps) F(u, v, w, x, y, z) = F 0 (u, v, x, z) + F 1 (u, w, y, z) = - PowerPoint PPT PresentationTRANSCRIPT
A Schema-Based Evolutionary Alg’m. for Black-Box Optimization
David A. CapeCS 448, Spring 2008
Missouri S & T
Motivation Arbitrary Additively Decomposable Functions
Example: multivariate polynomial (sum of two 4-bit D-Traps)
F(u, v, w, x, y, z) = F0(u, v, x, z) + F1(u, w, y, z) ={ 3[(1-u)(1-v)(1-x)(1-z)] + 2[u(1-v)(1-x)(1-z) + …]+ 1[uv(1-x)(1-z) + …] + 0[uvx(1-z) + …] + 4uvxz } +{ 3[(1-u)(1-w)(1-y)(1-z)] + 2[u(1-w)(1-y)(1-z) + …] + 1[uw(1-y)(1-z) + …] + 0[uwy(1-z) + …] + 4uwyz }= {5uvxz - u - v - x - z + 3} + {5uwyz - u - w - y - z + 3}
Building Block Hypothesis? F(1, 1, 1, 1, 1, 1) = 4+4 = 8 F(1, 1, 0, 1, 0, 1) = 4+1 = 5 F(1, 0, 1, 0, 1, 1) = 1+4 = 5 F(1, 0, 0, 0, 0, 1) = 1+1 = 2 F(1, 1, 0, 0, 0, 1) = 0+1 = 1 F(1, 1, 1, 1, 0, 1) = 4+0 = 4 Favg(1, #, #, #, #, 1) = [8+5+5+2+4(1)+4(4)] / 16 = 2.5 Favg(1, 1, #, #, #, 1) = [8+5+1+3(4)+2(0)] / 8 = 3.25 Favg(1, 1, #, 1, #, 1) = [8+5+2(4)] / 4 = 5.25 Favg(1, 1, 1, 1, #, 1) = [8+4)] / 2 = 6
Related Work Model-Building EAs use Estimation
of Distribution (EDA) techniques hBOA
Non-Model-Building EAs LLGA mGA TGA
Methodology Goals: Simplicity, generality, efficiency
“Don’t Care” symbols (#) as alleles Mutation from zero or one to # Mutation from # to zero or one Uniform crossover
Nondeterministic Representation Sampling of phenotypes for evaluation Small penalty for each # allele
“Agnostic EA” (AgEA) Allows ambiguity for each gene
Derived from schema theory
Uses traditional GA (TGA) operators
Duality between monomials and schemata
Experimental Design “Arbitrary additively decomposable”
Random multivariate polynomials Sums of trap subfunctions
Not necessarily concatenated Not necessarily adjacent
mGA with default parameters AgEA with equal number of
evaluations
AgEA vs. TGA on polynomials
Time to Find Best Fitness
0
10
20
30
40
50
0 2 4 6 8
Problem Difficulty
Avera
ge G
en
era
tio
ns
AgEA
TGA
(Problem difficulty was assessed subjectively)
Conclusion
Novel EA concept based on # alleles
Performs well on some simple problems
Better than competent EAs? hBOA?
Future Work
Comparison to messy GA, LLGA, hBOA
Careful analysis of data
Rigorous statistical tests
Meta-schema theory?
Questions?