1 / 17

Genetic Algorithms

Explore the advanced artificial intelligence technique of genetic algorithms for learning based on simulated evolution. Genetic algorithms begin with an initial population of hypotheses that undergo genetic operations to create the next generation. This robust method is easily parallelized, allowing for the creation of complex hypotheses with interacting parts. Dive into representing hypotheses with bit-strings, genetic operators, fitness functions, and selection processes. Discover how genetic algorithms evolve populations, tackle hypothesis space search challenges, and discuss the evolution of complete computer programs using genetic programming. Gain insights into models of evolution and learning, including the Lamarckian vs. Darwinian debate and the Baldwin Effect. Learn about parallelizing genetic algorithms through coarse and fine-grained parallel implementations.

kpye
Download Presentation

Genetic Algorithms

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Genetic Algorithms ML 9 Kristie Simpson CS536: Advanced Artificial Intelligence Montana State University

  2. Overview • Learning based on simulated evolution. • Begins with initial population of hypotheses. • Population uses genetic operations to create the next generation. • Most fit hypotheses survive.

  3. 9.1 Motivation • Successful, robust method for biological systems. • Hypotheses can contain complex interacting parts, where the impact of each part may be difficult to model. • Easily parallelized. • Easily understood.

  4. Terminology • Population – collection of hypotheses. • Fitness – numerical measure of the strength of a hypothesis. • Crossover (recombination) – two or more hypotheses combine to create one or more offspring. • Mutation – random modifications to individual hypotheses.

  5. 9.2 Genetic Algorithms • Table 9.1 pg. 251 • Population of initial hypotheses. • Fitness of each hypothesis is determined. • Create new generation. • Most fit hypotheses selected. • Crossover combines hypotheses to create offspring. • Mutation modifies individual hypotheses. • Update population.

  6. Where have we seen this before? • ACO 2.4.6 (pg. 55-57) Evolutionary Computation (Other Metaheuristics) • ACO 3.7.1 (pg. 93-99) Lamarckian vs. Darwinian (ACO plus Local Search) • ML 3.6.2 (pg. 65-66) Occam’s razor (Inductive Bias)

  7. 9.2.1 Representing Hypotheses • Hypotheses in GAs are often represented by bit-strings. IF Wind = Strong THEN PlayTennis = yes Outlook Wind PlayTennis 111 10 10

  8. 9.2.2 Genetic Operators

  9. 9.2.3 Fitness Function and Selection fitness proportionate selection: http://en.wikipedia.org/wiki/Fitness_proportionate_selection

  10. 9.3 An Illustrative Example • GABIL uses a GA to learn boolean concepts represented by a disjunctive set of propositional rules. • Hypotheses represented by bit-strings which grow with the number of rules. • Variable length bit-strings requires modification to the crossover rule.

  11. GAs for the TSP • http://www.ads.tuwien.ac.at/raidl/tspga/TSPGA.html

  12. 9.4 Hypothesis Space Search • GAs do not move smoothly from hypothesis to hypothesis (like Backpropagation). • Instead, they move much more abruptly and are less likely to fall into local minima. • Problem: crowding - highly fit individuals take over population. • Solution: alter the selection function (tournament, rank, fitness sharing, subspecies)

  13. 9.4.1 Population Evolution and the Schema Theorem • Mathematically characterize evolution. • Schemas – patterns that describe sets of bit strings (0s, 1s, *’s). • Evolution depends on selection, recombination, and mutation.

  14. 9.5 Genetic Programming • Extends genetic algorithms to the evolution of complete computer programs. • Population consists of computer programs rather than bit-strings. • Population of hypotheses typically represented by parse trees. • Fitness determined by executing the program on training data. • Crossover performed by swapping subtrees.

  15. 9.6 Models of Evolution and Learning • What is the relationship between learning during the lifetime of a single individual, and the longer time frame species-level learning afforded by evolution? • Lamarckian evolution – experiences of a single organism directly affect the genetic makeup of their offspring. • Scientific evidence contradicts this model.

  16. 9.6.2 Baldwin Effect • Evolution favors individuals with the capability to learn. • Individuals who learn rely less strongly on their genetic code. • Individual learning supports more rapid evolutionary progress, thereby increasing the chance that the species will evolve genetic, non-learned traits.

  17. 9.7 Parallelizing Genetic Algorithms • GAs naturally suited to parallel implementation. • Coarse grain – subdivide population into groups of individuals (demes). • Migration – individuals from one deme are copied/transferred to other demes. • Fine grain – one processor per individual in the population. • Recombination occurs among neighboring individuals.

More Related