1 / 9

Project 7 Genetic algorithms CS 539

Project 7 Genetic algorithms CS 539. Skyler Whorton November 27, 2012. Data and code. Project 3 datasets revisited Spambase for benchmarking and development purposes (No CFS) Faces for additional results (using CFS  56 pixels ) Genetic Algorithm Neural Network “GANN”

iokina
Download Presentation

Project 7 Genetic algorithms CS 539

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. Project 7GeneticalgorithmsCS 539 SkylerWhortonNovember 27, 2012

  2. Data and code • Project 3 datasets revisited • Spambase for benchmarking and development purposes (No CFS) • Faces for additional results (using CFS  56 pixels) • Genetic Algorithm Neural Network “GANN” • Original extensions to Weka’sMultilayerPerceptronclassifier • Static, user-specified hidden layer topology • Encode and evolve whole network of weights as an individual • Random initial weights as in MultilayerPerceptron • Fitness judged by classification accuracy over training data • Select the half of population with least error for reproduction • Crossover between networks; mutate weights individually, at random • Terminate after G generations, e.g. G=500

  3. Genetic Multilayer Perceptron • (Fixed) hypothesis lengthis a function of network topology • “Genes” are individual weights; “Chromosomes” are networks: • (0.1, 0.5, -0.3, -0.8, … 0.2, -0.7, 0.05, 0.0, 1.2, -0.9) • Initialize new network weights randomly within [-1.0, 1.0] • Generate initial population, P of network “representations” • Fitness: error rate over training data; no backpropogation • Select top 50% of most fit individuals in population i1 0.1 0.5 0.05 o1 h1 -0.3 i2 0.0 F -0.8 … 1.2 h2 o2 0.2 -0.9 im -0.7

  4. Crossover • Sample a pair of representations at a time. • Find random split point resulting in at least 10% exchange • Create and add two offspring from parts of parents. 0.1 0.5 -0.3 -0.8 0.2 -0.7 0.05 0.0 1.2 -0.9 Parents 0.2 -0.1 -0.2 0.6 0.15 -0.2 0.8 0.1 0.3 1.8 0.1 0.5 -0.3 0.6 0.15 -0.2 0.8 0.1 0.3 1.8 Offspring 0.2 -0.1 -0.2 -0.8 0.2 -0.7 0.05 0.0 1.2 -0.9

  5. Mutation • Mutate 10% of individual genes in all offspring, in-place • Apply mutation operators with uniform probability: • Double — W ← W * 2 • Halve — W ← W / 2 • Square— W ← W * W (removed) • Times Ten — W ← W * 10(removed) • Add Random— W ← W + Random(-1.0, 1.0) 0.1 0.5 -0.3 0.6 0.15 -0.2 0.8 0.1 0.3 1.8 Offspring Halve 0.1 0.5 -0.3 0.6 0.15 -0.2 0.4 0.1 0.3 1.8 Mutated

  6. GANN Variations - TRAINING

  7. ANN vs. GANN

  8. Genetic Programming • Goal: evolve individuals that represent entire programs • Experiment: learn an “interest” function: • A: accumulated moneyincluding interest • P: starting principal money[100, 1000000] • r: interest rate, [0.0, 1.0] • n: number of times tocompound yearly, [1, 12] * P Pow + n 1 / r n Needs to generate and learn a constant value

  9. Genetic Programming • Weka function “GeneticProgramming” classifier (3rd party) • Dataset: generate 1,000 “perfect” random tuples:221269, 0.25, 8, 280296.95(P) (r) (n) (A, computed value) • Options: 1000 generations, limit operators to +, *, /, ^ • Best (sadly) genetic program: vs. LinearRegression: / 1.646 * principal + 707902 * rate + 8224 * compounding + -409163 + 1 2.6 X 105 P RRSE: 56.39% RRSE: 21.51%

More Related