Line Search 2 Lecture 4 ME EN 575 Andrew Ning aning@byu.edu - - PDF document

line search 2
SMART_READER_LITE
LIVE PREVIEW

Line Search 2 Lecture 4 ME EN 575 Andrew Ning aning@byu.edu - - PDF document

Line Search 2 Lecture 4 ME EN 575 Andrew Ning aning@byu.edu Outline Root Finding Methods 1D Optimization Methods Root Finding Methods Root Finding Methods How do we know when we have reached a local minimum? Bisection Example:


slide-1
SLIDE 1

Line Search 2

Lecture 4

ME EN 575 Andrew Ning aning@byu.edu

Outline

Root Finding Methods 1D Optimization Methods

slide-2
SLIDE 2

Root Finding Methods Root Finding Methods

How do we know when we have reached a local minimum?

slide-3
SLIDE 3

Bisection Example: Refrigeration Tank

Minimize the cost of a cylindrical refrigeration tank with a volume of 50 m3.

  • Circular ends cost $10 per m2
  • Cylindrical walls cost $6 per m2
  • Refrigerator costs $80 per m2 over its life
slide-4
SLIDE 4

minimize 45πd2 + 17200 d with respect to d subject to d ≥ 0

Newton’s Method

We can do better by using gradient information

slide-5
SLIDE 5

1 2 3 x f(x)

xk+1 = xk − f ′(xk) f ′′(xk)

slide-6
SLIDE 6

f(x) x 1 2 3 x f(x) 2 1

Brent’s Method

slide-7
SLIDE 7

1D Optimization Methods Golden Section Search

slide-8
SLIDE 8

Extreme 1: bisection-like Extreme 2: small improvement

slide-9
SLIDE 9

I1 I2 I2 I2 = τI1 I1 I2 I2 I3 I3 I3 = τI2

slide-10
SLIDE 10

I1 = I2 + I3

Polynomial Methods

Approximate function locally as a polynomial (in this case quadratic): ˜ f = 1 2ax2 + bx + c If a > 0, the minimum of this function is x∗ = −b/a.

slide-11
SLIDE 11

Brent’s Method

Combines quadratic polynomial method with golden section search.

Where are we going?

Optimization N-D Optimization 1-D Optimization

  • Bisection Search
  • Fibonnacci Search
  • Golden Section Search
  • Newton’s Method
  • Polynomial Interpolation
  • Brent’s Method

Unconstrained Constrained

  • Lagrange Multipliers
  • Exterior Penalty Methods
  • Interior Point Methods
  • SQP

Line-search Methods

  • Steepest Descent
  • Conjugate Gradient
  • Newton’s Method
  • Quasi-Newton Methods

Smooth Non-smooth

  • Nelder-Mead Simplex
  • Genetic Algorithms

Trust-region Methods

  • Polynomial Fits
  • Nonparametric Fits
slide-12
SLIDE 12

Intuition in Higher Dimensions

Consider a hypersphere inscribed inside a hypercube volume of sphere ? volume of cube

1 2 3 4 5 6 7 8 9 dimension 0.0 0.2 0.4 0.6 0.8 1.0 Vsphere/Vcube