SPSA Algorithm.rarSPSA Algorithm

function [theta,x] =SPSA(x0,fun,a,c,alpha,gamma,A,n) theta = x0; p = length(x0); for k = 0:n-1 ak = a/(k+1+A)^alpha; ck = c/(k+1)^gamma; delta = 2*round(rand(1,p))-1; thetaplus = theta+ck*delta; thetaminus = theta-ck*delta; yplus = fun(thetaplus); yminus = fun(thetaminus); ghat = (yplus-yminus)./(2*ck*delta); x(k+1,:) = theta'; theta = theta-ak*ghat; end %Notes: %If maximum and minimum values on the values of theta can be %specified, say thetamax and thetamin, then the following two %lines can be added below the theta update line to impose the %constraints % theta=min(theta,thetamax); % theta=max(theta,thetamin); %The MATLAB feval operation (not used above) is useful in %yplus and yminus evaluations to allow for easy change of %loss function. %Algorithm initialization not shown above (see discussion %in introduction to MATLAB code).