Rapid Accurate Optimization of Difficult Problems Using Fast Messy Genetic Algorithms D.E. Goldberg, K. Deb, H. Kargupta & G. Harik, 1993 Presented by : Yann SEMET Visiting Student Universite de Technologie de Compiegne France
Roadmap & Foreword Framework & motivations Brief review of mGAs What ’s new ? Probabilistically Complete Initialization Building Block Filtering Thresholding revisited Experiments & Analysis Discussion
Framework & Motivations From simple GAs to fmGAs : Design of competent GAs Tackling hard,deceptive problems How : Taking care of the building blocks Messiness : mGAs Efficiency : fmGAs
Messy GAs ? mGAs are different from simple GAs : variable string length no fixed locus cut & splice Over and under specification Primordial and juxtapositional phases Inner, outer loops and k-wise processing
Complexity concerns 1 single copy of  all  k-substrings !
Toward Fast Processing Avoid the initialization Bottleneck partial enumeration = unGAlike ? Back to earlier unsuccessful ideas : probabilistic initialization shortening the strings 2 suggested improvements : probabilistically complete initialization and BB filtering.
Probabilistically Complete Initialization Random instead of explicit l’=l-k is a good choice Overall Cost : O(l)
Building Block Filtering Reducing string length down to k By performing alternatively : Selection alone Random deletion Don ’t delete mindlessly ! Key : more duplication than deletion
Building Block Filtering 2
Overall Complexity Initialization : O(l) BB filtering : O(log(l)) Overall :O(l.log(l)) Fits the remainder of the mGA! Goal achieved.
Thresholding revisited Restrict competition among too dissimilar BBs Genic Thresholding mechanism New values :
Experiments Concatenation of k-trap functions No mutation p(splice)=1 ; p(cut)=0.03 Competitive template : 000… l ’=l-k No k-wise processing
Base-Line results k=3 26650 function calls instead of 40600
Larger scale problems K=5 ; 10, 14 and 30 times 1.9*10^5 function calls for 150-bits !
Analysis & Strengths Winning heuristics Conservative assumptions No parameter fiddling Theoretical basis The long-short-long signature
Summary 1 problem : complexity 2 solutions : stochasticity and filtering 1 result : overall cost O(l.log(l)) Successful experiments
Conclusions Natural Fuzziness is good Yet, supports the Design Decomposition//BBs approach
Future Challenges 1 Other misleading factors : Mixed scale and size problems Crosstalk Massive multimodality
Future Challenges 2 Suggested Carry-ons : Floating point codes Classifier codes Permutation codes (OmeGA)
Discussion and Critique Close to the real world ? Deceptive functions and real problems Conservative assumptions The natural metaphor Messiness or decomposition ? The long-short-long signature

Fm G As

  • 1.
    Rapid Accurate Optimizationof Difficult Problems Using Fast Messy Genetic Algorithms D.E. Goldberg, K. Deb, H. Kargupta & G. Harik, 1993 Presented by : Yann SEMET Visiting Student Universite de Technologie de Compiegne France
  • 2.
    Roadmap & ForewordFramework & motivations Brief review of mGAs What ’s new ? Probabilistically Complete Initialization Building Block Filtering Thresholding revisited Experiments & Analysis Discussion
  • 3.
    Framework & MotivationsFrom simple GAs to fmGAs : Design of competent GAs Tackling hard,deceptive problems How : Taking care of the building blocks Messiness : mGAs Efficiency : fmGAs
  • 4.
    Messy GAs ?mGAs are different from simple GAs : variable string length no fixed locus cut & splice Over and under specification Primordial and juxtapositional phases Inner, outer loops and k-wise processing
  • 5.
    Complexity concerns 1single copy of all k-substrings !
  • 6.
    Toward Fast ProcessingAvoid the initialization Bottleneck partial enumeration = unGAlike ? Back to earlier unsuccessful ideas : probabilistic initialization shortening the strings 2 suggested improvements : probabilistically complete initialization and BB filtering.
  • 7.
    Probabilistically Complete InitializationRandom instead of explicit l’=l-k is a good choice Overall Cost : O(l)
  • 8.
    Building Block FilteringReducing string length down to k By performing alternatively : Selection alone Random deletion Don ’t delete mindlessly ! Key : more duplication than deletion
  • 9.
  • 10.
    Overall Complexity Initialization: O(l) BB filtering : O(log(l)) Overall :O(l.log(l)) Fits the remainder of the mGA! Goal achieved.
  • 11.
    Thresholding revisited Restrictcompetition among too dissimilar BBs Genic Thresholding mechanism New values :
  • 12.
    Experiments Concatenation ofk-trap functions No mutation p(splice)=1 ; p(cut)=0.03 Competitive template : 000… l ’=l-k No k-wise processing
  • 13.
    Base-Line results k=326650 function calls instead of 40600
  • 14.
    Larger scale problemsK=5 ; 10, 14 and 30 times 1.9*10^5 function calls for 150-bits !
  • 15.
    Analysis & StrengthsWinning heuristics Conservative assumptions No parameter fiddling Theoretical basis The long-short-long signature
  • 16.
    Summary 1 problem: complexity 2 solutions : stochasticity and filtering 1 result : overall cost O(l.log(l)) Successful experiments
  • 17.
    Conclusions Natural Fuzzinessis good Yet, supports the Design Decomposition//BBs approach
  • 18.
    Future Challenges 1Other misleading factors : Mixed scale and size problems Crosstalk Massive multimodality
  • 19.
    Future Challenges 2Suggested Carry-ons : Floating point codes Classifier codes Permutation codes (OmeGA)
  • 20.
    Discussion and CritiqueClose to the real world ? Deceptive functions and real problems Conservative assumptions The natural metaphor Messiness or decomposition ? The long-short-long signature