Evolutionary computation and_applications

52
Evolutionary Computation And Applications Deepak Dawar 03/15/22 DD - North Dakota State University 1

Transcript of Evolutionary computation and_applications

Evolutionary Computation And

ApplicationsDeepak Dawar

04/15/23 DD - North Dakota State University 1

Problems Galore!

204/15/23 DD - North Dakota State University

The Traveling Salesman

• NP Complete• It is likely that the worst case

running time for any algorithm for the TSP increases exponentially with the number of cities.

304/15/23 DD - North Dakota State University

Evolutionary Algorithms

404/15/23 DD - North Dakota State University

Classes of Search Methods

Search Method

Calculus Based Method

Guided Random Search Method

Enumerative Method

BFSDFS Dynamic Programming

Tabu Search

Hill Climbing

Simulated Anealing

Evolutionary Algorithms

Genetic Algorithms

Fibonacci Sort

Swarm Intelligence

Particle Swarm

Ant Colony

Differential Evolution

04/15/23 DD - North Dakota State University 5

Evolutionary Algorithms

• Emulate evolutionary processes.• Operate on population of individuals.

604/15/23 DD - North Dakota State University

List of Parts for Real Parameter Optimization with EA

• A function to optimize

• A domain in which to optimize

• An Evolutionary Algorithm

704/15/23 DD - North Dakota State University

The Function and The Domain

Function:We want to find the minimum value on the landscape.

8

Global Minimum

04/15/23 DD - North Dakota State University

The Algorithm-Differential Evolution

• Real Parameter Optimizer• Utilizes a pre-specified number of random solutions • Continuously improves them through a series of mutations and re-

combinations• Usually, the number of pre-specified solutions i.e. the population size

does not change during the lifetime of the algorithm

904/15/23 DD - North Dakota State University

Differential Evolution

1004/15/23 DD - North Dakota State University

Important Nomenclature

• D = Dimensionality of the problem• NP = No. of solution vectors in a

population • G = No. of generations the population has to go

through• F = Scaling Factor• Cr = Crossover rate

1104/15/23 DD - North Dakota State University

Initialization

1204/15/23 DD - North Dakota State University

Initial Population

1304/15/23 DD - North Dakota State University

Mutation

• Mutation is a change in the gene characteristics of a chromosome.• Applied to evolutionary computation it means a change in the

parameters of the vector through a perturbation with a random element.

1404/15/23 DD - North Dakota State University

Mutation

• A parent vector from the current generation is called target vector• The mutant vector obtained through differential perturbation is

called donor vector • The offspring obtained through recombination of target and donor is

called trial vector

1504/15/23 DD - North Dakota State University

Mutation

1604/15/23 DD - North Dakota State University

Mutation

1704/15/23 DD - North Dakota State University

Crossover

1804/15/23 DD - North Dakota State University

Create Trial Vector

• Suppose: rand =0.945, Cr =0.9

• If rand < Cr, pick the parameter from the donor

• Else from the parent

• Do this for all 10 parameters

1904/15/23 DD - North Dakota State University

Selection- Choose between Parent and Child/Trial Vector

2004/15/23 DD - North Dakota State University

Control Parameters

• NP, F and Cr are collectively termed as control parameters of DE• The performance of DE is very sensitive to the values of control

parameters• Every problem/function may respond differently to different sets of

control parameters.

2104/15/23 DD - North Dakota State University

Benchmark Functions- IEEE CEC 2013 Test Suite

• State of the art benchmark functions used in CEC Real Parameter Optimization Competitions and Conferences

• Emulate the properties of real world large scale optimization problems

• IEEE CEC Test suites have constantly evolved over time with the advances in the field of Large Scale Global Optimization

• All the problems listed in the Test Suite are minimization problems

2204/15/23 DD - North Dakota State University

Rotated Ackley’s Function

2304/15/23 DD - North Dakota State University

Scaling Factor

Scales the difference of solutions and controls the step size

Many Authors have reported that its effective values lie between [0.4, 1]

though it may ultimately also depend upon the problem landscape

Contradicting results have been reported for what should be good values

for the scaling factor

So what do we do?

04/15/23 DD - North Dakota State University 24

Altering the Scaling Vector

Deterministic - the parameters are altered based on some user defined rules

Adaptive - the parameters are allowed to adapt based on some feedback from

the algorithm

Evolutionary - the parameters are encoded into the solution itself and they

evolve as a part of the general population

04/15/23 DD - North Dakota State University 25

In Our First Paper

• We primarily focus our attention towards deterministic parameter control methods and the control parameter, scale factor (F), in particular.

04/15/23 DD - North Dakota State University 26

Dither

04/15/23 DD - North Dakota State University 27

DETVSF - DE with Time Varying Scale Factor

Linear Reduction

04/15/23 DD - North Dakota State University 28

DEDASF - DE with Dither and Annealed Scale Factor

Non Linear Reduction

04/15/23 DD - North Dakota State University 29

Birds Eye View

04/15/23 DD - North Dakota State University 30

What do we achieve?

04/15/23 DD - North Dakota State University 31

Result 1 - FriedMan’s Test - Ranks

04/15/23 DD - North Dakota State University 32

Result 2 - Hocheberg’s Post Hoc Procedure: significance level 0.1

04/15/23 DD - North Dakota State University 33

Second Paper - Objective

• To detect, classify and count the type of moving vehicles

34Source: http://www.123rf.com/stock-photo/highway.html04/15/23 DD - North Dakota State University

Vehicle Recognition Steps

3504/15/23 DD - North Dakota State University

How to Identify Axles?

• Axles are circular

• Shape recognition algorithm must be employed

3604/15/23 DD - North Dakota State University

Hough Transform

• Is a feature extraction technique used in image analysis• Can be used to detect lines, circles, and ellipses or any shape

represented by a set of parameters• We focus on circle detection

3704/15/23 DD - North Dakota State University

Recognition Example (1)

• Original Image

3804/15/23 DD - North Dakota State University

Recognition Example (1)

• dp (accumulator resolution)=1.25

3904/15/23 DD - North Dakota State University

Recognition Example (2)

4004/15/23 DD - North Dakota State University

Recognition Example (3)

4104/15/23 DD - North Dakota State University

04/15/23 DD - North Dakota State University 42

DEMO

Paper 3 - Object Tracking

04/15/23 DD - North Dakota State University 43

Stages In Object Tracking

• Object Initialization• Appearance model Initialization• Search – Object Localization• Appearance model Updation

04/15/23 DD - North Dakota State University 44

Modeling The Target

• Represent the target as• Vector• Matrix• Histogram• Wavelets

04/15/23 DD - North Dakota State University 45

Modeling – A Difficult Task

• Scene changes• Illumination • Background• Occlusion• Deformation• Noise

04/15/23 DD - North Dakota State University 46

Univariate Gaussian

04/15/23 DD - North Dakota State University 47

Multivariate Gaussian

04/15/23 DD - North Dakota State University 48

Gaussian Mixture

04/15/23 DD - North Dakota State University 49

Segmentation using GMM

04/15/23 DD - North Dakota State University 50

Demo

04/15/23 DD - North Dakota State University 51

04/15/23 DD - North Dakota State University 52

Thank You