function?[thetaPe]=nnekf（thetaPxyQR）
%?NNEKF?????A?function?using?the?EKF?to?training?a?MLP?NN
%?[thetaPz]=nnekf（thetaPxyQR）?searches?the?optimal?parameters?theta?
%?of??a?MLP?NN?based?on?a?set?of?training?data?with?input?x?and?output?y.
%?Input:
%???theta:?initial?guess?of?MLP?NN?parameter.?The?network?structure?is
%???determined?by?the?number?of?parameters?ns?the?number?of?inputs?（size?of
%???x）nx?and?the?number?of?output?（size?of?y）?ny.?The?euqation?of?the?NN
%???is:?y?=?W2?*?tanh（?W1?*?x?+?b1）?+?b2?and?theta?=?[W1（:）;b1;W2（:）;b2].
%???Therefore?ns?=?nx?*?nh?+?nh?+?nh?*?ny?+?ny?which?gives?the?number?of
%???hidden?nodes?is?nh?=?（ns?-?ny）?/?（nx?+?ny?+?1）;
%???P:?the?covariance?of?the?initial?theta.?Needs?to?be?tuned?to?get?good
%???training?performance.
%???x?and?y:?input?and?output?data?for?training.?For?batch?training?x?and
%???y?should?be?arranged?in?such?a?way?that?each?observation?corresponds?to?
%???a?column.
%???Q:?the?virtual?process?covariance?for?theta?normally?set?to?very?small
%???values.
%???R:?the?measurement?covariance?dependen?on?the?noise?level?of?data?tunable.?
%
%?Example:?a?NN?model?to?approximate?the?sin?function
%?rand（‘state‘0）
%?N=20;
%?Ns=100;
%?x=1.2*randn（NNs）;
%?y=sin（x）+0.1*randn（NNs）;
%?z=y;
%?nh=4;
%?ns=nh*2+nh+1;
%?theta=randn（ns1）;
%?P=diag（[100*ones（1nh*2）?10000*ones（1nh+1）]）;
%?Q=0.0