Genetic Algorithm and Other Experimental Design Methods

Embed Size (px)

Citation preview

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    1/29

    06/27/2011

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    2/29

    Search Techniqes

    Calculus Base

    TechniqesGuided random search

    techniques

    Enumerative

    Techniqes

    BFSDFS Dynamic

    Programming

    Tabu Search Hill

    Climbing

    Simulated

    Anealing

    Evolutionary

    Algorithms

    Genetic

    Programming

    Genetic

    Algorithms

    Fibonacci Sort

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    3/29

    The Genetic Algorithm

    y Originally developed by John Holland (1975)

    y Aclass of probabilistic optimization and search

    algorithms based on the mechanics of naturalselection and genetic inheritance.

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    4/29

    Quick Overviewy Inspired bynatural evolution

    y Population of individuals

    y Individual is feasible solution to problem

    y

    Each individual is characterized by a Fitness functiony Higher fitness is better solution

    y Based on their fitness, parents are selected to reproduceoffspring for a new generation

    y Fitter individuals have more chance to reproduce

    y New generation has same size as old generation; old generationdies

    y Offspring has combination of properties of two parents

    y If well designed, population will converge to optimal solution

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    5/29

    Components of a GAAproblem to solve, and ...

    y Encoding technique (gene, chromosome)

    y Initialization procedure (creation)y Evaluation function (environment)

    y Selection of parents (reproduction)

    y Genetic operators (mutation, Crossover)

    y Termination Criteria (End GA)

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    6/29

    Encoding in GAy Binary Encoding

    In binary encoding every chromosome is a string ofbits, 0 or 1

    y Value Encoding In value encoding, every chromosome is a string of some values.

    Values can be anything connected to problem, real numbers or

    chars to some complicated objects.

    y Tree Encoding In tree encoding every chromosome is a tree of some objects, such

    as functions or commands in programming language.

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    7/29

    Selection in GA

    y Roulette wheel selection

    y Rank Selection

    y Tournament Selection

    y Stochastic universal sampling

    y SteadyState Selection

    y Elitism

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    8/29

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    9/29

    Mutation Generating new offspring from single parent

    y Maintaining the diversity of the individualsy

    Crossover can only explore the combinations of the currentgene pool

    y Mutation can generate new genes

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    10/29

    A Simple Example[1]y Simple problem: max x2 over {0,1,,31}

    y GAapproach:

    y Representation: binary code, e.g. 01101m 13y Population size: 4

    y 1-point xover, bitwise mutation

    y Roulette wheel selection

    y Random initialization

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    11/29

    x2

    example: selection

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    12/29

    X2 example: crossover

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    13/29

    X2 example: mutation

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    14/29

    Genetic Programming

    Objective: Find a computer program with inputvariables (X1,X2,,Xn) whose output

    equals the given data

    1 Terminalset: T = {X1,X2,..,Xn, Random-

    Constants}

    2 Function set: F = {+, -, *, %...etc.}

    3 Fitness: The sum of the absolute value of the

    difference between the candidate

    programs output and the given data

    (computed over numerous values of the

    variables)

    4 Parameters: Population sizeM

    5 Termination: An individualemerges whosesum of

    absoluteerrors isless than 0.1

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    15/29

    Using Trees To Represent Computer

    Programs(+ 2 3 (* X7) (/ Y5))

    +

    2 3 * /

    X 7 5Y

    Functions

    Terminals

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    16/29

    Multi-Objective Optimization

    y Uses concept of pareto-optimality.

    y Investigate a set of solutions, each of which satisfies

    the objectives at an acceptable level without beingdominated by any other solution.

    y Diversity of solutions are maintained by fitnesssharing and niching.

    y Elitism is used.

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    17/29

    NSGA II[2]

    y Aset of N population is chosen. Set t=0, P0y Q0 is generated by first selecting solutions from P0 and then

    doing Crossover and Mutationy Rt = Pt UQty Identification of non dominated fronts F1,.Fk in Rt.y

    For i=1,kCalculate crowding distance of the solution in Fi.

    Pt+1 is created as follows : -Case I If [Pt+1] + [Fi] N , then add least crowded

    N-[Pt+1] solutions from Fi to Pt+1 Use binary tournament selection based on crowding distance to

    select parents from Pt+1

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    18/29

    When to Use a GAy Alternate solutions are too slow or overly complicated

    y Need an exploratory tool to examine new approaches

    y Problem is similar to one that has already beensuccessfully solved by using a GA

    y Want to hybridize with an existing solution

    y Benefits of the GAtechnology meet key problem

    requirements

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    19/29

    Some GA Application Types

    Domain Application Types

    Control gas pipeline, pole balancing, missile evasion, pursuit

    Design semiconductor layout, aircraft design, keyboardconfiguration, communication networks

    Schedulingmanufacturing, facility scheduling, resource allocation

    Robotics trajectory planning

    Machine Learning designing neural networks, improving classificationalgorithms, classifier systems

    Signal Processing filter design

    Game Playing poker, checkers, prisoners dilemma

    CombinatorialOptimization

    set covering, travelling salesman, routing, bin packing,graph colouring and partitioning

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    20/29

    Benefits of Genetic Algorithms

    y Concept is easy to understand

    y Modular, separate from application

    y Supports multi-objective optimizationy Good for noisy environments

    y Always an answer; answer gets better with time

    y Inherently parallel; easily distributed

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    21/29

    TaguchiMethod Developed byDr. Genichi Taguchi.

    Investigates how different parameters affect the meanand variance of a process performance characteristics.

    Involves use oforthogonal arrays. Analysis of variance on the collected data from theTaguchi design of experiments can be used to selectnew parameter values to optimize the performance

    characteristics.

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    22/29

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    23/29

    Taguchis Orthogonal Array Selector

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    24/29

    Analysis of Experimental Data

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    25/29

    Analysis of Experimental Data

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    26/29

    Limitationsy Interactions are part of the real world. In Taguchi's

    arrays, interactions are confounded and difficult to

    resolve.y Relative method, do not exactly indicate the true

    effects of parameters on performance characteristicsvalue.

    y

    Offline method, so inappropriate for a dynamicallychanging process, such as simulation study.

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    27/29

    Response Surface Methodology

    y Developed by Box andWilson in 1950.

    y RSM usually involves three stages:

    (1) Design of experiments

    (2)Response surface modeling through regression

    (3) Optimization

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    28/29

    Steps ofRSMThe steps by which the response surface method should beundertaken can be summarized as follows:

    y Conduct the experiment with the independent variables varyingaround the present operating point.

    y Obtain a fitted equation with data obtained in the experiment.

    Normally, regression methods are used in this step. Frequently, alinear model represents the model sufficiently well.

    y Move the experimental point in the direction of steepest ascent(or descent if a minimum is sought) and repeat the previoussteps.

    y When little improvement is obtained, the optimum is near.y Conduct a 3-level factorial experiment around this point.

    y Obtain a fitted quadratic equation by regression methods.

    y Based on this quadratic equation, determine the optimum

  • 8/6/2019 Genetic Algorithm and Other Experimental Design Methods

    29/29

    Referencesy [1] GeneticAlgorithms by David E. Goldberg, 1989.y [2] Multi-Objective optimization using genetic

    algorithm byA. Konak, D. Coit,A. Smith,2006.