what are NARX Function inputs "X" and "Xi"? Whats is an example of both?

Allow me to preface I am somewhat new to Matlab, and Neural Networks. Despite this I have created a NARX function to predict multiple steps ahead, It has been trained and I am happy with the output.

1. Why is the output of shorter length than the input? I understand there is a shifting timeframe, so does this mean once the delay is removed, I am getting the prediction 1 step ahead?...hence the difference length?

2. What form do the inputs to the function take. I can see that X is a cell array of 'input' and 'target', but what is "Xi" and "-"

Whare are 'input delay states' and why a 2x2 cell?

% Xi = 2x2 cell 2, initial 2 input delay states. % Each Xi{1,ts} = 1xQ matrix, initial states for input #1. % Each Xi{2,ts} = 1xQ matrix, initial states for input #2.

function [Y,Xf,Af] = myNeuralNetworkFunction(X,Xi,~) %MYNEURALNETWORKFUNCTION neural network simulation function. % % Generated by Neural Network Toolbox function genFunction, 11-Aug-2017 08:47:44. % % [Y,Xf,Af] = myNeuralNetworkFunction(X,Xi,~) takes these arguments: % % X = 2xTS cell, 2 inputs over TS timesteps % Each X{1,ts} = 1xQ matrix, input #1 at timestep ts. % Each X{2,ts} = 1xQ matrix, input #2 at timestep ts. % % Xi = 2x2 cell 2, initial 2 input delay states. % Each Xi{1,ts} = 1xQ matrix, initial states for input #1. % Each Xi{2,ts} = 1xQ matrix, initial states for input #2. % % Ai = 2x0 cell 2, initial 2 layer delay states. % Each Ai{1,ts} = 10xQ matrix, initial states for layer #1. % Each Ai{2,ts} = 1xQ matrix, initial states for layer #2. % % and returns: % Y = 1xTS cell of 2 outputs over TS timesteps. % Each Y{1,ts} = 1xQ matrix, output #1 at timestep ts. % % Xf = 2x2 cell 2, final 2 input delay states. % Each Xf{1,ts} = 1xQ matrix, final states for input #1. % Each Xf{2,ts} = 1xQ matrix, final states for input #2. % % Af = 2x0 cell 2, final 0 layer delay states. % Each Af{1ts} = 10xQ matrix, final states for layer #1. % Each Af{2ts} = 1xQ matrix, final states for layer #2. % % where Q is number of samples (or series) and TS is the number of timesteps.

%#ok<*RPMT0>

% ===== NEURAL NETWORK CONSTANTS =====

% Input 1 x1_step1.xoffset = 1.6758; x1_step1.gain = 2.56574727389352; x1_step1.ymin = -1;

% Input 2 x2_step1.xoffset = 1.6758; x2_step1.gain = 2.56574727389352; x2_step1.ymin = -1;

% Layer 1 b1 = [2.5713526297760567196;-2.0096071730491433804;1.219887714138954804;1.5780200416537291108;0.25816093584503307934;0.18074593510612815828;0.70718796247587500936;-0.64421851569067067889;-1.7185815490127793748;2.6543854248661524764]; IW1_1 = [-0.16869542244623322857 -1.1712370140185244249;1.7602248106014910523 1.7561838897145030103;-1.4532469319368439553 -0.20396299675279816466;-1.2259478848587126443 1.442053061998331609;1.0740243238755720068 -1.496438098993799537;0.11179379980948041251 -1.7099632172394532148;-0.013198512334931664786 0.38902590336003639582;-0.33160470518089074643 -0.87688059602713563923;-2.2476574460726266302 -1.0059096087535042141;0.38300707211150541998 -0.2317876417318727178]; IW1_2 = [1.7933907521546201824 1.0452899104755479787;0.45472529234613878746 1.9311747950676203534;-1.2824630670549146405 -0.65160240846466466191;-1.1203828086453961888 -0.67877992281461829727;0.019576524613524378532 -1.5007152873252320724;0.9449462904564566168 0.98542655430178127673;-0.58499269871995085435 2.0323103368093251575;1.4512761139696179757 1.5427372349806167673;1.1522033707836289995 -0.36687974450126964454;1.3760674937190295886 -1.9352496628660595945];

% Layer 2 b2 = -0.27118710073177765274; LW2_1 = [0.66967935811621492892 0.12797614436784843228 0.43538693605978834311 0.046209366565424604689 0.15444805117195750666 0.53450744499729629933 0.11465064833947560818 1.0837510139972506007 -0.15469602738558474453 -0.41093430476120473838];

% Output 1 y1_step1.ymin = -1; y1_step1.gain = 2.56574727389352; y1_step1.xoffset = 1.6758;

% ===== SIMULATION ========

% Format Input Arguments isCellX = iscell(X); if ~isCellX X = {X}; end if (nargin < 2), error('Initial input states Xi argument needed.'); end

% Dimensions TS = size(X,2); % timesteps if ~isempty(X) Q = size(X{1},2); % samples/series elseif ~isempty(Xi) Q = size(Xi{1},2); else Q = 0; end

% Input 1 Delay States Xd1 = cell(1,3); for ts=1:2 Xd1{ts} = mapminmax_apply(Xi{1,ts},x1_step1); end

% Input 2 Delay States Xd2 = cell(1,3); for ts=1:2 Xd2{ts} = mapminmax_apply(Xi{2,ts},x2_step1); end

% Allocate Outputs Y = cell(1,TS);

% Time loop for ts=1:TS

      % Rotating delay state position
      xdts = mod(ts+1,3)+1;

    % Input 1
    Xd1{xdts} = mapminmax_apply(X{1,ts},x1_step1);

    % Input 2
    Xd2{xdts} = mapminmax_apply(X{2,ts},x2_step1);

    % Layer 1
    tapdelay1 = cat(1,Xd1{mod(xdts-[1 2]-1,3)+1});
    tapdelay2 = cat(1,Xd2{mod(xdts-[1 2]-1,3)+1});
    a1 = tansig_apply(repmat(b1,1,Q) + IW1_1*tapdelay1 + IW1_2*tapdelay2);

    % Layer 2
    a2 = repmat(b2,1,Q) + LW2_1*a1;

    % Output 1
    Y{1,ts} = mapminmax_reverse(a2,y1_step1);
end

% Final Delay States finalxts = TS+(1: 2); xits = finalxts(finalxts<=2); xts = finalxts(finalxts>2)-2; Xf = [Xi(:,xits) X(:,xts)]; Af = cell(2,0);

% Format Output Arguments if ~isCellX Y = cell2mat(Y); end end

% ===== MODULE FUNCTIONS ========

% Map Minimum and Maximum Input Processing Function function y = mapminmax_apply(x,settings) y = bsxfun(@minus,x,settings.xoffset); y = bsxfun(@times,y,settings.gain); y = bsxfun(@plus,y,settings.ymin); end

% Sigmoid Symmetric Transfer Function function a = tansig_apply(n,~) a = 2 ./ (1 + exp(-2*n)) - 1; end

% Map Minimum and Maximum Output Reverse-Processing Function function x = mapminmax_reverse(y,settings) x = bsxfun(@minus,y,settings.ymin); x = bsxfun(@rdivide,x,settings.gain); x = bsxfun(@plus,x,settings.xoffset); end

Answer 1

[Faiz Gouri]

The following documents will be helpful for you-

Design Time Series NARX Feedback Neural Network

preparets

[Greg Heath]

You might find some of my posts in both NEWSGROUP and ANSWERS helpful

 greg narxnet

Hope this helps.

Greg