<- previous    index    next ->

Lecture 17, Optimization, finding minima

A difficult numerical problem is the finding of a global minima
of an unknown function. This type of problem is called an
optimization problem because the global minima is typically
the optimum solution.

In general we are given a numerically computable function of
some number of parameters  v = f(x_1, x_2, ... , x_n)
and must find the values of x_1, x_2, ... , x_n that gives
the smallest value of v. Or, by taking the absolute value,
find the values of x_1, x_2, ... , x_n that give the value
of v closest to zero.

Generally the problem is bounded and there are given maximum and
minimum values for each parameter. There are typically many
places where local minima exists. Thus, the general solution
must include a global search then a local search to find the
local minima. There is no general guaranteed optimal solution.

First consider a case of only one variable on a non differentiable
function, y = f(x) where x has bounds xmin and xmax. There may be
many local minima, valleys that are not the deepest.

  *     *       *      *         *
   *   * *     * *    *  *      * *   *
    * *   *   *   *  *  * *    *   * * 
     *     * *     **      *  *     *
            *                *
                            *

Global search:
Do a y=f(x) for x = xmin, x = x+dx, until x>xmax,
Save the smallest y and the x0=x for that y.

Then find the local minimum:
Consider evaluating f(x) at some initial point x0 and x0+dx.
If f(x0) < f(x0+dx) you might move to x0-dx.
if f(x0) > f(x0+dx) you might move to x0+dx+dx.
The above may be very bad choices!

Here are the cases you should consider:
Compute  yl=f(x-dx)  y=f(x)  yh=f(x+dx)  for some dx

      yh      y            yh       y      yl         yl
    y           yh    yl         yl             yh       y
 yl        yl            y            yh      y            yh

 case 1    case 2     case 3     case 4    case 5     case6

For your next three points, always keep best x:
case 1 x=x-dx                possibly dx=2*dx
case 2 x=x-dx  dx=dx/2 
case 3 dx=dx/2
case 4 x=x+dx  dx=dx/2
case 5 dx=dx/2
case 6 x=x+dx                possibly dx=2*dx
Then loop.

A simple version in Python is
optm.py
optm_py.out
optm.py running
minimum (x-5)**4 in -10 to +10 initial 2, 0.001, 0.001, 200
xbest= 5.007 , fval= 2.401e-09 ,cnt= 21
 
minimum (x-5)**4 in -10 to +10 initial 2 using default
xbest= 4.998271 , fval= 8.9367574925e-12 ,cnt= 35

Another version, found on the web, slightly modified
min_search.py
min_search_py.out
min of (x-5)**4, h= 0.1
xmin= 4.9984375
fmin= 5.96046447754e-12
n= 86 , sx= 4.996875 , xx= 5.003125
 
min of (x-5)**4, h= 0.01   # smaller initial h, not as good
xmin= 4.9975
fmin= 3.90625000037e-11
n= 704 , sx= 4.995 , xx= 5.005
 
min of (x-5)**4, h= 2.0    # smaller tolerance, better
xmin= 4.99993896484
fmin= 1.38777878078e-17
n= 47 , sx= 4.99987792969 , xx= 5.00012207031

 

Beware local minima

Both of the above, will find the best of a local minima. There could be a local minima, thus when dx gets small enough, remember the best x and use another global search value to look for a better optimum. Some heuristics may be needed to increase dx. This is one of many possible algorithms. Another algorithm that is useful for large areas in two dimensions for z=f(x,y) is: Use a small dx and dy to evaluate a preferred direction. Use an expanding search, doubling dx and dy until no more progress is made. Then use a contracting search, halving dx and dy to find the local minima on that direction. Repeat until the dx and dy are small enough. The numbers indicate a possible order of evaluation of the points (in one dimension). 1 2 3 5 6 7 Value computed 4 ^ 8 | 9 ------------> X direction Numbers are sample number. Note dx doubles for samples 2, 3, 4, 5 then dx is halved and added to best so far, 4, to get 6 then halved to get 7, 8, 9. There may be many expansions and contractions. The pseudo derivatives are used to find the preferred direction: (After finding the best case from above, make positive dx and dy best.) z=f(x,y) zx=f(x+dx,y) zx < z zy=f(x,y+dy) zy < z r=sqrt(((z-zx)^2+(z-zy)^2)) dx=dx*(z-zx)/r dy=dy*(z-zy)/r This method has worked well on the spiral trough. The really tough problems have many discontinuities. I demonstrated a function that was everywhere discontinuous. The function was f(x)=x^2-1 with f(x)=1 if the bottom bit of x^2-1 is a one. discontinuous.c first function discontinuous.out Another, recursive, function that is continuous, yet in the limit of recursing, nowhere differentiable, is: double f(x){ /* needs a termination condition, e.g count=40 */ if(x<=1.0/3.0) return (3.0/4.0)*f(3.0*x); else if(x>=2.0/3.0) return (1.0/4.0)+(3.0/4.0)*f(3.0*x-2.0); else /* 1/3 < x < 2/3 */ return (1.0/4.0)+(1.0/2.0)*f(2.0-3.0*x);} discontinuous.c second function discontinuous.out In general, it will not work to use derivatives, or even pseudo derivatives without additional heuristics. A sample program that works for some functions of three floating point parameters is shown below. Then, a more general program with a variable number of parameters is presented with a companion crude global search program. Two parameter optimization in Python useful for project optm2.py3 source code optm2_py3.out output global_search.py3 source code search_py3.out output search.py3 source code search2_py3.out output notice that zbest stops getting better at some small dx, dy mpmath_example.py3 source code mpmath_example_py3.out output test_mpmath.py3 source code test_mpmath_py3.out output Two parameter optimization in java global_search.java source code global_search_java.out output search2.java source code search2_java.out output notice that zbest stops getting better at some small dx, dy Three parameter optimization: optm3.h optm3.c test_optm3.c test_optm3_c.out N parameter optimization in C and java: This includes a global search routine srchn. optmn.h optmn.c test_optmn.c test_optmn_c.out optmn_dev.java test_optmn_dev.java test_optmn_dev_java.out optmn_interface.java optmn_function.java user modifies optmn.java optmn_run.java optmn_run_java.out In MatLab use "fminsearch" see the help file. Each search is from one staring point. You need nested loops to try many starting points. I got error warnings that I ignored, OK to leave them in your output. optm.m An interesting test case is a spiral trough: Possibly a better view: test_spiral.f90 test_spiral_f90.out spiral.f90 test_spiral.c test_spiral_c.out spiral.h spiral.c test_spiral.java test_spiral_java.out spiral_trough.py test_spiral.py test_spiral_py.out Your project is a multiple precision minimization See Lecture 15 for multiple precision
    <- previous    index    next ->

Other links

Go to top