Genetic algorithms and Machine Learning

D. GoldbergJ. Holland

article1988Machine Learning3,445 citations

Explains how genetic algorithms and classifier systems use implicit parallelism and evolutionary mechanisms to efficiently search complex spaces and discover reusable building blocks for machine learning.

Listen

This 1988 guest editorial argues that genetic algorithms offer machine learning a robust, nature-inspired method for searching complex spaces by mimicking evolutionary processes of selection, recombination, and building-block assembly. The authors, David Goldberg and John Holland, address the question of why learning systems should draw from evolution rather than solely from brain-like models, noting that evolution has produced highly complex adaptations over time despite its apparent slowness in natural settings. They emphasize that artificial systems can compress these timescales dramatically while retaining the same mechanisms for handling novelty and incremental improvement.

The editorial sets out to explain the core properties of genetic algorithms and classifier systems, to counter common objections to the evolutionary metaphor, and to frame the papers selected for a special double issue of the journal. It draws on prior theoretical work, including schema theorems that establish implicit parallelism, and on early applications to illustrate how populations of string-encoded solutions can evaluate and exploit useful substrings far more efficiently than explicit enumeration. The discussion covers six representative papers that range from medical image registration under noise to scaling classifier systems on parallel hardware and comparing credit-assignment methods.

The central claims are that genetic algorithms achieve both explicit parallelism through population-based sampling and substantial implicit parallelism by processing many more component patterns than the population size; that classifier systems allow new rules to be inserted and tested without disrupting existing performance; and that reproduction plus recombination enables rapid development of appropriate complexity across a wide range of problems. The authors further note that these systems require no global consistency checks and can operate incrementally, making them suitable for environments that exhibit perpetual novelty where conventional search methods are likely to fail.

These properties imply that genetics-based approaches can reduce the cost and risk of exploring large, poorly understood design spaces while supporting graceful integration of learned and programmed knowledge. The editorial observes that early commercial uses already existed by the late 1980s, suggesting that the methods had moved beyond theory into practical deployment. It positions the special issue as evidence that the field had advanced enough to warrant broader attention from the machine-learning community.

Further work should focus on less traditional languages with simpler syntax that can be manipulated reliably by genetic operators, on tighter integration of symbolic and subsymbolic representations, and on empirical comparisons across more domains. The authors point to the 1985 and 1987 International Conferences on Genetic Algorithms as sources of additional breadth, from VLSI layout to automated generation of LISP code. Because the piece is an editorial rather than a controlled study, its claims rest on the cited theoretical results and the range of papers presented; readers should treat the performance expectations as directional guidance pending larger-scale validation on contemporary hardware and problem sizes.

Goldberg et al (1988).pdf
Cover for Genetic algorithms and Machine Learning

Abstract

This document is a guest editorial and does not contain an abstract.

Table of Contents

  • Genetic Algorithms and Machine Learning
  • Metaphors for learning
  • Genetic algorithms and classifier systems
  • Arguments for the evolutionary metaphor
  • Contents of the special issue
  • References

Knowls

  1. Knowl 1 — Genetic Algorithms and Implicit Parallelism

    model/method

    Genetic algorithms (GAs) are probabilistic, population-based search and optimization procedures designed to operate on large search spaces where states are represented as strings. In addition to explicit parallelism derived from simultaneously sampling a population of mm candidate solutions, genetic algorithms exhibit implicit parallelism: processing a population of mm strings implicitly evaluates and biases the search across substantially more than m3m^3 component substrings (schemata). Above-average substrings serve as building blocks that are recombined to exploit regularities in complex, noisy, or poorly understood problem spaces.

  2. Knowl 2 — Classifier Systems Architecture and Conflict Resolution

    model/method

    A classifier system is a genetics-based parallel production system structured to exploit the implicit parallelism of genetic algorithms. Its core architectural properties comprise:

    1. Standardized Message Passing: All communication between rules (classifiers) and environmental interfaces occurs via standardized messages. Rule conditions are defined by the messages they match, and rule actions are defined by the messages they post.
    2. Computational Completeness and Uniform Syntax: The system provides a computationally complete execution environment whose regular, string-based syntax allows genetic operators to manipulate, discover, and recombine candidate rules as building blocks.
    3. Competition-Based Conflict Resolution and Gracefulness: Conflicts between active rules are resolved dynamically via competition rather than explicit global consistency algorithms. Consequently, new candidate rules and trial hypotheses can be introduced incrementally without corrupting previously acquired capabilities.
  3. Knowl 3 — Computational Viability of the Evolutionary Metaphor in Machine Learning

    theoretical result

    The argument that evolutionary mechanisms are fundamentally too slow for artificial learning—based on the billions of years required by biological evolution—is refuted by two principles:

    1. Time-Scale Disparity: Artificial computational cycles operate orders of magnitude faster than biological reproductive generations.
    2. Recombination of Building Blocks: Efficient adaptation does not proceed via naive selection and point mutation alone. The combination of reproduction, recombination (crossover), and building-block processing under implicit parallelism enables the rapid synthesis of complex, high-performing structures across broad domains without sacrificing search generality.

Coverage note — Summaries and quotations of the individual papers included in the special double issue (authored by Fitzpatrick and Grefenstette, De Jong, Robertson and Riolo, Booker, Belew and Forrest, and Grefenstette) were omitted as they constitute the original contributions of those referenced papers.

References

  1. 1.Bateson, G. (1972). Steps to an ecology of mind. New York: Ballantine.
  2. 2.Davis, L., & Coombs, S. (1987). Genetic algorithms and communication link speed design: Theoretical considerations. Genetic Algorithms and Their Applications: Proceedings of the Second International Conference on Genetic Algorithms (pp. 252-256). Cambridge, MA: Lawrence Erlbaum.
  3. 3.Edelman, G. M. (1987). Neural Darwinism: The theory of neuronal group selection. New York: Basic Books.
  4. 4.Fourman, M. P. (1985). Compaction of symbolic layout using genetic algorithms. Proceedings of the First International Conference on Genetic Algorithms and Their Applications (pp. 141-152). Pittsburgh, PA: Lawrence Erlbaum.
  5. 5.Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and machine learning. Reading, MA: Addison-Wesley.
  6. 6.Grefenstette, J. J. (Ed.). (1985). Proceedings of the First International Conference on Genetic Algorithms and Their Applications. Pittsburgh, PA: Lawrence Erlbaum.
  7. 7.Grefenstette, J. J. (Ed.). (1987). Genetic Algorithms and Their Applications: Proceedings of the Second International Conference on Genetic Algorithms. Cambridge, MA: Lawrence Erlbaum.
  8. 8.Holland, J. H. (1962). Outline for a logical theory of adaptive systems. Journal of the Association for Computing Machinery, 3, 297-314.
  9. 9.Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor, MI: University of Michigan Press.
  10. 10.Holland, J. H., Holyoak, K. J., Nisbett, R. E., & Thagard, P. R. (1986). Induction: Processes of inference, learning, and discovery. Cambridge, MA: MIT Press.
  11. 11.Minsky, M. (1986). The society of mind. New York: Simon and Schuster.
  12. 12.Waddington, C. H. (1967). Summary discussion. In P. S. Moorhead & M. M Kaplan (Eds.), Mathematical challenges to the neo-Darwinian interpretation of evolution. Philadelphia, PA: Wistar Institute Press.

Citation

MLA
Goldberg, D., and J. Holland. “Genetic Algorithms and Machine Learning”. Machine Learning, vol. 3, nos. 2-3, 1988, pp. 95–99, https://doi.org/10.1007/BF00113892.
APA
Goldberg, D., & Holland, J. (1988). Genetic algorithms and Machine Learning. Machine Learning, 3(2-3), 95–99. https://doi.org/10.1007/BF00113892
Chicago
Goldberg, D., and J. Holland. 1988. “Genetic Algorithms and Machine Learning”. Machine Learning 3 (2-3): 95–99. https://doi.org/10.1007/BF00113892.
Harvard
Goldberg, D. and Holland, J. (1988) “Genetic algorithms and Machine Learning”, Machine Learning, 3(2-3), pp. 95–99. Available at: https://doi.org/10.1007/BF00113892.
Vancouver
1. Goldberg D, Holland J (1988) Genetic algorithms and Machine Learning. Machine Learning 3:95–99

BibTeX

@article{goldberg1988genetic,
  title = {Genetic algorithms and Machine Learning},
  author = {Goldberg, D. and Holland, J.},
  year = {1988},
  journal = {Machine Learning},
  publisher = {Springer Science and Business Media LLC},
  volume = {3},
  number = {2-3},
  pages = {95-99},
  doi = {10.1007/BF00113892},
  url = {http://dx.doi.org/10.1007/BF00113892}
}
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF