optimize odes in parallel -凯发k8网页登录

optimize odes in parallel

this example shows how to optimize parameters of an ode.

it also shows how to avoid computing the objective and nonlinear constraint function twice when the ode solution returns both. the example compares patternsearch and ga in terms of time to run the solver and the quality of the solutions.

you need a parallel computing toolbox™ license to use parallel computing.

step 1. define the problem.

the problem is to change the position and angle of a cannon to fire a projectile as far as possible beyond a wall. the cannon has a muzzle velocity of 300 m/s. the wall is 20 m high. if the cannon is too close to the wall, it has to fire at too steep an angle, and the projectile does not travel far enough. if the cannon is too far from the wall, the projectile does not travel far enough either.

air resistance slows the projectile. the resisting force is proportional to the square of the velocity, with proportionality constant 0.02. gravity acts on the projectile, accelerating it downward with constant 9.81 m/s². therefore, the equations of motion for the trajectory x(t) are

$\frac{d^{2} x (t)}{d t^{2}} = - 0.02 ‖ v (t) ‖ v (t) - (0, 9.81),$

where $v (t) = d x (t) / d t .$

the initial position x0 and initial velocity xp0 are 2-d vectors. however, the initial height x0(2) is 0, so the initial position depends only on the scalar x0(1). and the initial velocity xp0 has magnitude 300 (the muzzle velocity), so depends only on the initial angle, a scalar. for an initial angle th, xp0 = 300*(cos(th),sin(th)). therefore, the optimization problem depends only on two scalars, so it is a 2-d problem. use the horizontal distance and the angle as the decision variables.

step 2. formulate the ode model.

ode solvers require you to formulate your model as a first-order system. augment the trajectory vector (x₁(t),x₂(t)) with its time derivative (x'₁(t),x'₂(t)) to form a 4-d trajectory vector. in terms of this augmented vector, the differential equation becomes

$\frac{d}{d t} x (t) = [\begin{matrix} x_{3} (t) \\ x_{4} (t) \\ - .02 ‖ (x_{3} (t), x_{4} (t)) ‖ x_{3} (t) \\ - .02 ‖ (x_{3} (t), x_{4} (t)) ‖ x_{4} (t) - 9.81 \end{matrix}] .$

write the differential equation as a function file, and save it on your matlab^® path.

function f = cannonfodder(t,x)
f = [x(3);x(4);x(3);x(4)]; % initial, gets f(1) and f(2) correct
nrm = norm(x(3:4)) * .02; % norm of the velocity times constant
f(3) = -x(3)*nrm; % horizontal acceleration
f(4) = -x(4)*nrm - 9.81; % vertical acceleration

visualize the solution of the ode starting 30 m from the wall at an angle of pi/3.

step 3. solve using patternsearch.

the problem is to find initial position x0(1) and initial angle x0(2) to maximize the distance from the wall the projectile lands. because this is a maximization problem, minimize the negative of the distance (see ).

to use patternsearch to solve this problem, you must provide the objective, constraint, initial guess, and options.

these two files are the objective and constraint functions. copy them to a folder on your matlab path.

function f = cannonobjective(x)
x0 = [x(1);0;300*cos(x(2));300*sin(x(2))];
sol = ode45(@cannonfodder,[0,15],x0);
% find the time t when y_2(t) = 0
zerofnd = fzero(@(r)deval(sol,r,2),[sol.x(2),sol.x(end)]);
% then find the x-position at that time
f = deval(sol,zerofnd,1);
f = -f; % take negative of distance for maximization
function [c,ceq] = cannonconstraint(x)
ceq = [];
x0 = [x(1);0;300*cos(x(2));300*sin(x(2))];
sol = ode45(@cannonfodder,[0,15],x0);
if sol.y(1,end) <= 0 % projectile never reaches wall
    c = 20 - sol.y(2,end);
else
    % find when the projectile crosses x = 0
    zerofnd = fzero(@(r)deval(sol,r,1),[sol.x(2),sol.x(end)]);
    % then find the height there, and subtract from 20
    c = 20 - deval(sol,zerofnd,2);
end

notice that the objective and constraint functions set their input variable x0 to a 4-d initial point for the ode solver. the ode solver does not stop if the projectile hits the wall. instead, the constraint function simply becomes positive, indicating an infeasible initial value.

the initial position x0(1) cannot be above 0, and it is futile to have it be below –200. (it should be near –20 because, with no air resistance, the longest trajectory would start at –20 at an angle pi/4.) similarly, the initial angle x0(2) cannot be below 0, and cannot be above pi/2. set bounds slightly away from these initial values:

lb = [-200;0.05];
ub = [-1;pi/2-.05];
x0 = [-30,pi/3]; % initial guess

set the usecompletepoll option to true. this gives a higher-quality solution, and enables direct comparison with parallel processing, because computing in parallel requires this setting.

opts = optimoptions('patternsearch','usecompletepoll',true);

call patternsearch to solve the problem.

tic % time the solution
[xsolution,distance,eflag,outpt] = patternsearch(@cannonobjective,x0,...
    [],[],[],[],lb,ub,@cannonconstraint,opts)
toc

optimization terminated: mesh size less than options.meshtolerance
 and constraint violation is less than options.constrainttolerance.
xsolution =
  -28.8123    0.6095
distance =
 -125.9880
eflag =
     1
outpt = 
  struct with fields:
         function: @cannonobjective
      problemtype: 'nonlinearconstr'
       pollmethod: 'gpspositivebasis2n'
    maxconstraint: 0
     searchmethod: []
       iterations: 5
        funccount: 269
         meshsize: 8.9125e-07
         rngstate: [1×1 struct]
          message: 'optimization terminated: mesh size less than options.meshtolerance↵ and constraint violation is less than options.constrainttolerance.'
elapsed time is 0.792152 seconds.

starting the projectile about 29 m from the wall at an angle 0.6095 radian results in the farthest distance, about 126 m. the reported distance is negative because the objective function is the negative of the distance to the wall.

visualize the solution.

x0 = [xsolution(1);0;300*cos(xsolution(2));300*sin(xsolution(2))];
sol = ode45(@cannonfodder,[0,15],x0);
% find the time when the projectile lands
zerofnd = fzero(@(r)deval(sol,r,2),[sol.x(2),sol.x(end)]);
t = linspace(0,zerofnd); % equal times for plot
xs = deval(sol,t,1); % interpolated x values
ys = deval(sol,t,2); % interpolated y values
plot(xs,ys)
hold on
plot([0,0],[0,20],'k') % draw the wall
xlabel('horizontal distance')
ylabel('trajectory height')
legend('trajectory','wall','location','nw')
ylim([0 70])
hold off

step 4. avoid calling the expensive subroutine twice.

both the objective and nonlinear constraint function call the ode solver to calculate their values. use the technique in to avoid calling the solver twice. the runcannon file implements this technique. copy this file to a folder on your matlab path.

function [x,f,eflag,outpt] = runcannon(x0,opts)
if nargin == 1 % no options supplied
    opts = [];
end
xlast = []; % last place ode solver was called
sol = []; % ode solution structure
fun = @objfun; % the objective function, nested below
cfun = @constr; % the constraint function, nested below
lb = [-200;0.05];
ub = [-1;pi/2-.05];
% call patternsearch
[x,f,eflag,outpt] = patternsearch(fun,x0,[],[],[],[],lb,ub,cfun,opts);
    function y = objfun(x)
        if ~isequal(x,xlast) % check if computation is necessary
            x0 = [x(1);0;300*cos(x(2));300*sin(x(2))];
            sol = ode45(@cannonfodder,[0,15],x0);
            xlast = x;
        end
        % now compute objective function
        % first find when the projectile hits the ground
        zerofnd = fzero(@(r)deval(sol,r,2),[sol.x(2),sol.x(end)]);
        % then compute the x-position at that time
        y = deval(sol,zerofnd,1);
        y = -y; % take negative of distance
    end
    function [c,ceq] = constr(x)
       ceq = [];
        if ~isequal(x,xlast) % check if computation is necessary
            x0 = [x(1);0;300*cos(x(2));300*sin(x(2))];
            sol = ode45(@cannonfodder,[0,15],x0);
            xlast = x;
        end
        % now compute constraint functions
        % first find when the projectile crosses x = 0
        zerofnd = fzero(@(r)deval(sol,r,1),[sol.x(1),sol.x(end)]);
        % then find the height there, and subtract from 20
        c = 20 - deval(sol,zerofnd,2);
    end
end

reinitialize the problem and time the call to runcannon.

x0 = [-30;pi/3];
tic
[xsolution,distance,eflag,outpt] = runcannon(x0,opts);
toc

optimization terminated: mesh size less than options.meshtolerance
 and constraint violation is less than options.constrainttolerance.
elapsed time is 0.670715 seconds.

the solver ran faster than before. if you examine the solution, you see that the output is identical.

step 5. compute in parallel.

try to save more time by computing in parallel. begin by opening a parallel pool of workers.

parpool

starting parpool using the 'local' profile ...
connected to the parallel pool (number of workers: 6).
ans = 
 processpool with properties: 
            connected: true
           numworkers: 6
              cluster: local
        attachedfiles: {}
    autoaddclientpath: true
          idletimeout: 30 minutes (30 minutes remaining)
          spmdenabled: true

set the options to use parallel computing, and rerun the solver.

opts = optimoptions('patternsearch',opts,'useparallel',true);
x0 = [-30;pi/3];
tic
[xsolution,distance,eflag,outpt] = runcannon(x0,opts);
toc

optimization terminated: mesh size less than options.meshtolerance
 and constraint violation is less than options.constrainttolerance.
elapsed time is 1.894515 seconds.

in this case, parallel computing was slower. if you examine the solution, you see that the output is identical.

step 6. compare with the genetic algorithm.

you can also try to solve the problem using the genetic algorithm. however, the genetic algorithm is usually slower and less reliable.

as explained in , you cannot save time when using ga by the nested function technique used by patternsearch in step 4. instead, call ga in parallel by setting the appropriate options.

options = optimoptions('ga','useparallel',true);
rng default % for reproducibility
tic % time the solution
[xsolution,distance,eflag,outpt] = ga(@cannonobjective,2,...
    [],[],[],[],lb,ub,@cannonconstraint,options)
toc

optimization terminated: average change in the fitness value less than options.functiontolerance
 and constraint violation is less than options.constrainttolerance.
xsolution =
  -37.6217    0.4926
distance =
 -122.2181
eflag =
     1
outpt = 
  struct with fields:
      problemtype: 'nonlinearconstr'
         rngstate: [1×1 struct]
      generations: 4
        funccount: 9874
          message: 'optimization terminated: average change in the fitness value less than options.functiontolerance↵ and constraint violation is less than options.constrainttolerance.'
    maxconstraint: 0
       hybridflag: []
elapsed time is 12.529131 seconds.

the ga solution is not as good as the patternsearch solution: 122 m versus 126 m. ga takes more time: about 12 s versus under 2 s; patternsearch takes even less time in serial and nested, less than 1 s. running ga serially takes even longer, about 30 s in one test run.

optimize odes in parallel -凯发k8网页登录