Upload
scot-stone
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
March 23, 2003 AAAI symposium, Stanford.
Jianjun Hu, Erik D. Goodman, Kisung Seo Zhun Fan, Ronald C. Rosenberg
Genetic Algorithm Research & Applications Group (GARAGe)
Michigan State University
HFC: a Continuing EA Framework for Scalable Evolutionary Synthesis
March 23, 2003 AAAI symposium, Stanford.
Evolutionary Synthesis: the problem
Topology innovation: How many and what types of building blocks, How to connect them?
Parameter innovation: How to size the elements?
R
S
C I
R1
Input:
>Building blocks>Function Specification(Evaluation function)>EA settings (operators, parameters…)
Output:
Design solutions
Evolution
March 23, 2003 AAAI symposium, Stanford.
Sustainable & scalable evolutionary synthesis: Definition and Reality
Definition: the capability to obtain better results or solve larger scale problems when given more computing resourcesReality: not sustainable and not scalable
Bloating and aging problem (e.g.:Innovation in 50 generations in GP)Demand for huge population size (GP)Premature convergence and local optima (all EA)
But biological evolution is sustainable and scalable!
March 23, 2003 AAAI symposium, Stanford.
Sustainable & scalable evolutionary synthesis: two types of obstacles
Two types of obstacles:1. Convergent nature of current EA framework
one of major factors leading to :GP aging phenomenonpremature convergencedependence on huge population size
2. Non-scalable compositional mechanisms for topological and parametric evolution
possible solutions: modularity, hierarchical organization, biological developmental principles…
This paper: addresses the EA convergence problem
March 23, 2003 AAAI symposium, Stanford.
Comparison with biological evolution: how to achieve sustainable evolution
Biological evolutionAlmost infinite population sizeSimultaneous evolution of all levels of organisms: bacteria coexist with humans Fair competition: different levels of organism coexist in different nichesSustainable innovation possible
Artificial evolution (EA)Limited population sizeFocusing on current high-fitness individuals Unfair competition: highly-evolved individuals compete with new offspring of low fitnessInnovation capability rapidly depleted
March 23, 2003 AAAI symposium, Stanford.
Sustainable evolution: example & comparisonSustainable education system: how generations of talents are educated?
Unsustainable EA: how generations of solutions are evolved?
Academiclevel
primaryschool
middleschool
highschool
college
graduateschool
New babi es
Simultaneoustraining at alllevels
Fitness(usually also
complexity in GP)
High-fitnessindividuals
Randomindividuals
One-epochsearch
Gen1
Gen10
Gen100
Gen1000
Gen50000
March 23, 2003 AAAI symposium, Stanford.
Assembly-line structured continuing EAs
randomindividuals
individuals containingintermediate levels of
building blocks
end products: thesolution
building blocks from low order tohigh order
Lowlevel
Highlevel
Fitness from lower to higher
March 23, 2003 AAAI symposium, Stanford.
HFC-EA framework
(a) In HFC model, subpopulations are organized in ahierarchy with ascending fitness level. Each level (with oneor more subpopulations) accomodates individuals within acertain fitness range determined by the admission thresholds
fitness
fmin
fmaxsubpop9
subpop8
subpop7
subpop5,6
subpop3,4
subpop0,1,2
ADT5
AdmissionBuffers
randomindividualgenerator
ADT:Admission Threshold
ADT4
ADT3
ADT2
ADT1
(b) HFC model extends the search horizontally in searchspace and vertically in fitness dimension and kills badindividuals at appropriate times while allowing promisingyoung individuals grow up continuously
level 1
level 2
level 3
level 4
level 5
fitness
The Streamlined Structure of HFC Model (a) and HFCfrom the Perspective of Landscape (b).
level 6
March 23, 2003 AAAI symposium, Stanford.
System synthesis problem: eigenvalue placement
-2.0 -3.3 -2.0 3.3 -7.5 -4.5 -7.5 4.5 -3.5 -12.0 -3.5 12.0 -3.4 -12.0 -3.4 12.0-10.0 -8.0-10.0 8.0
-2.063 -3.005 -2.063 3.005 -7.205 -4.581 -7.205 4.581 -3.580 -11.835 -3.580 11.835 -3.334 -12.526 -3.334 12.526 -9.997 -8.040
-9.997 8.040 Max distance error 0.530
Average distance error 0.272
2 hours
Sf
1Se 1
1
0
C
R
I
1
R C
I
0
R C
C
0 RL
March 23, 2003 AAAI symposium, Stanford.
Experiment result: sustainability& robustnessDynamic system synthesis problem with simultaneous
topology and parameter search
Average Eigenvalue Location Error vs Evaluation NumberHFC-GP
Evaluation Number (thousands)
0 50 100 150 200 250 300 350
Loca
tion
Err
or
0.0
.5
1.0
1.5
2.0
STD
Average Eigenvalue Location Error vs Evaluation NumberMultiPopulation-GP
Evaluation Number (thousands)
0 50 100 150 200 250 300 350
Loca
tion
Err
or
0.0
.2
.4
.6
.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
STD
March 23, 2003 AAAI symposium, Stanford.
Experiment result: handling GP aging problem
0 200 400 600 800 1000 12000
0.2
0.4
0.6
0.8
1
generation
sta
ndard
fitness
Standardized fitness of Best Individual of run
OnePopMulPopHFC-GP
10 eigenvalue dynamic system synthesis problem
March 23, 2003 AAAI symposium, Stanford.
Experiment result: small population size works equally well
10-parity problem with function set {and, xor, or, not)
Success Rate
0%10%20%30%40%50%60%70%80%90%
100%
150 250 400 800PopSize
succ
ess
rate
OnePop
MulPop
HFC
HFC_ADM
HFCATP
March 23, 2003 AAAI symposium, Stanford.
Why HFC works: the explanation
Fitness LevelsNormal EA individualsHFC individuals
Random Individual Generator
March 23, 2003 AAAI symposium, Stanford.
Conclusion: Hierarchical Fair Competition Principle for EA
HFC is very effective in evolutionary synthesisSimultaneous evolution at all (fitness) levels, from the random population to best individualsAvoid premature convergence by allowing emergence of new optima rather than trying to jump out of local optimaAllow use of strong selection pressure without risk of premature convergenceSmall population size also works
March 23, 2003 AAAI symposium, Stanford.
Ongoing & future work
Developing single population HFC (CHFC) (with continuous level segmentations) algorithm to achieve sustainable evolutionDeveloped HFC-enhanced multi-objective EAs (GECCO2003)To develop hybrid parallel-HFC GA/GP system where each deme is implemented as a CHFC populationTo develop multi-level system synthesis: from framework evolution to complete solutions
March 23, 2003 AAAI symposium, Stanford.
Generalization of HFC-EA Framework
A generic framework for continuing sustainable evolutionary computation (GA, GP, ES, …)Especially good for Evolutionary Synthesis for Sustainable topological innovation which has extremely rugged fitness landscape.Especially effective for problems with:
High multi-modality Strong tendency of premature convergenceRequirement on robustnessRequirement on adaptation in dynamic environment
Also applicable for artificial life evolution
March 23, 2003 AAAI symposium, Stanford.
Scaling Mechanism of HFC
A natural parallel evolutionary computation model. Better than island parallel modelHybridizing with single-population HFC EAs (Each deme is a sustainable HFC subpop)Natural hybridizing with explicit hierarchical building block discovery mechanismsAllow using small population size and longer time to achieve good resultsNo restart to waste computing effort