# Optimization methods Review

## Presentation on theme: "Optimization methods Review"— Presentation transcript:

Optimization methods Review
Mateusz Sztangret Faculty of Metal Engineering and Industrial Computer Science Department of Applied Computer Science and Modelling Krakow, r.

Outline of the presentation
Basic concepts of optimization Review of optimization methods gradientless methods, gradient methods, linear programming methods, non-deterministic methods Characteristics of selected methods method of steepest descent genetic algorithm

Basic concepts of optimization
Man’s longing for perfection finds expression in the theory of optimization. It studies how to describe and attain what is Best, once one knows how to measure and alter what is Good and Bad… Optimization theory encompasses the quantitative study of optima and methods for finding them. Beightler, Phillips, Wilde Foundations of Optimization

Basic concepts of optimization
Optimization /optimum/ - process of finding the best solution Usually the aim of the optimization is to find better solution than previous attached

Basic concepts of optimization
Specification of the optimization problem: definition of the objective function, selection of optimization variables, identification of constraints.

Mathematical definition
where: x is the vector of variables, also called unknowns or parameters; f is the objective function, a (scalar) function of x that we want to maximize or minimize; gi and hi are constraint functions, which are scalar functions of x that define certain equations and inequalities that the unknown vector x must satisfy.

Set of allowed solutions
Constrain functions define the set of allowed solution that is a set of points which we consider in the optimization process. X Xd

Solution is called global minimum if, for all
Obtained solution Solution is called global minimum if, for all Solution is called local minimum if there is a neighbourhood N of such that Global minimum as well as local minimum is never exact due to limited accuracy of numerical methods and round off error

Local and global solutions
f(x) local minimum global minimum x

Problems with multimodal objective function
f(x) start start x

Discontinuous objective function
f(x) Discontinuous function x 3

Minimum or maximum f(x) f c x* x – c – f

General optimization flowchart
Start Set starting point x(0) i = 0 i = i + 1 Calculate f(x(i)) NO Stop condition x(i+1) = x(i) + Δx(i) YES Stop

Commonly used stop conditions are as follows:
obtain sufficient solution, lack of progress, reach the maximum number of iterations

Classification of optimization methods
Classification of optimizing methods due to: The type of solved problem Linear programming Nonlinear optimization Constraints Optimization with constraints Optimization without constraints Size of the problem One-dimensional methods multidimensional methods The optimization criteria One-objective methods Multiobjective methods

The are several type of optimization algorithms: gradientless methods,
Optimization methods The are several type of optimization algorithms: gradientless methods, line search methods, multidimensional methods, gradient methods, linear programming methods, non-deterministic methods

Multidimensional methods
Gradientless methods Line search methods Expansion method Golden ratio method Multidimensional methods Fibonacci method Method based on Lagrange interpolation Hooke-Jeeves method Rosenbrock method Nelder-Mead simplex method Powell method

Advantages: simplicity, they do not require computing derivatives of the objective function. Disadvantages: they find first obtained minimum they demand unimodality and continuity of objective function

Method of steepest descent Conjugate gradients method Newton method
Gradient methods Method of steepest descent Conjugate gradients method Newton method Davidon-Fletcher-Powell method Broyden-Fletcher-Goldfarb-Shanno method

Advantages: simplicity, greater effciency in comparsion with gradientless methods. Disadvantages: they find first obtained minimum they demand unimodality, continuity and differentiability of objective function

Linear programming If both the objective function and constraints are linear we can use one of the linear programming method: Graphical method Simplex method

Non-deterministic method
Monte Carlo method Genetic algorithms Evolutionary algorithms strategy (1 + 1) strategy (μ + λ) strategy (μ, λ) Particle swarm optimization Simulated annealing method Ant colony optimization Artificial immune system

Features of non-deterministic methods
Advantages: any nature of optimised objective function, they do not require computing derivatives of the objective function. Disadvantages: high number of objective function calls

Optimization with constraints
Ways of integrating constrains External penalty function method Internal penalty function method

Multicriteria optimization
In some cases solved problem is defined by few objective function. Usually when we improve one the others get whose. weighted criteria method ideal point method

Weighted criteria method
Method involves the transformation multicriterial problem into one-criterial problem by adding particular objective functions.

Ideal point method In this method we choose an ideal solution which is
outside the set of allowed solution and the searching optimal solution inside the set of allowed solution which is closest the the ideal point. Distance we can measure using various metrics Ideal point Allowed solution

Method of steepest descent
Algorithm consists of following steps: Substitute data: u0 – starting point maxit – maximum number of iterations e – require accuracy of solution i = 0 – iteration number Compute gradient in ui

Method of steepest descent
Choose the search direction Find optimal solution along the chosen direction (using any line search method). If stop conditions are not satisfied increased i and go to step 2.

Zigzag effect Let’s consider a problem of finding minimum of function: f(u)=u12+3u22 Starting point: u0=[-2 3] Isolines

Algorithm consists of following steps:
Genetic algorithm Algorithm consists of following steps: Creation of a baseline population. Compute fitness of whole population Selection. Crossing. Mutation. If stop conditions are not satisfied go to step 2.

Creation of a baseline population
Genotype Objective function value (f(x)=x2) 28900 7225 44944 33124 1849 51984

Selection Baseline population Parents’ population

Roulette wheel method

Crossing crossing point Descendant individual no 1 0 1 0
Parent individual no 1 Parent individual no 2

Mutation Parent individual

Mutation r>pm r>pm r<pm Mutation

Mutation Mutation 1 0 0 0 1 0 1 0 r<pm r>pm r>pm r>pm

Mutation r<pm r>pm Mutation

Mutation Parent individual 1 0 1 0 1 0 1 0
Descendant individual

Genetic algorithm After mutation, completion individuals are recorded in the descendant population, which becomes the baseline population for the next algorithm iteration. If obtained solution satisfies stop condition procedure is terminated. Otherwise selection, crossing and mutation are repeated.