SLMC Home
LWPR tutorial
This page demonstrates the basic usage of our LWPR library based on the Matlab implementation. We'll fit an LWPR model to a simple one-dimensional toy dataset. This tutorial directly corresponds to the Matlab script "test_lwpr_1D.m". Please note that our library contains two further demonstration scripts for higher-dimensional data ("test_lwpr_2D.m" and "test_lwpr_nD.m"), which should be self-explanatory once you went through this tutorial.
% TEST_LWPR_1D % % Simple script to demonstrate the LWPR algorithm % We train a model on toy data, and later retrieve % predictions and confidence bounds from the model
Step 0: Generate training data
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % In the next lines, we create a toy dataset with Ntr=500 samples %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Ntr = 500;
Xtr = 10*rand(Ntr,1);
testfunc = inline('10*sin(7.8*log(1+x))./(1+0.1*x.^2)');
Ytr = 5 + testfunc(Xtr) + 0.1*randn(Ntr,1).*Xtr;
clf
plot(Xtr,Ytr,'r.');
drawnow;
hold on
A possible resulting plot, showing our training data, is depicted as follows. Note how the noise variance increases with the magnitude of the input data.
Step 1: Setup LWPR model, choose tuning parameters
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Now we create a new LWPR model using "lwpr_init" % The parameters (1,1) mean 1 input dimension, 1 output dimension % % Using "lwpr_set", we set appropriate values for model parameters % The most important parameter is "init_D", the initial distance % metric that is assigned to new receptive fields (local models). % In general init_D is a square matrix corresponding to the input % dimensionality. For 1-D it is just a scalar. % Smaller values for init_D yield wider receptive fields, which % might oversmooth at the start of training and lead to slow % convergence. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
model = lwpr_init(1,1); model = lwpr_set(model,'init_D',20); model = lwpr_set(model,'update_D',1); model = lwpr_set(model,'diag_only',1); model = lwpr_set(model,'penalty',0.0001); model = lwpr_set(model,'init_alpha', 40);
By default, LWPR uses Gaussian kernels as the activation function of all receptive fiels. You may alternatively employ the Bisquare kernel with the following line:
model = lwpr_set(model,'kernel','BiSquare');
Now, "model" is a Matlab data structure that contains the above settings, and default values for any other LWPR parameter.
>> model
model =
nIn: 1
nOut: 1
n_data: 0
mean_x: 0
var_x: 0
diag_only: 1
update_D: 1
meta: 0
meta_rate: 250
penalty: 1.0000e-04
init_alpha: 40
norm_in: 1
norm_out: 1
init_D: 20
init_M: 4.4721
name: ''
w_gen: 0.1000
w_prune: 1
init_lambda: 0.9990
final_lambda: 1.0000
tau_lambda: 0.9999
init_S2: 1.0000e-10
add_threshold: 0.5000
kernel: 'Gaussian'
sub: [1x1 struct]
Step 1a: Optional, but recommended speed-up
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % With the following line, we transfer the LWPR model (which is a % Matlab struct until now) into MEX-internal memory. This drastically % speeds up predictions und updates, but has the slight disadvantage % that its parameters can not be directly accessed anymore. % After the call to "lwpr_storage", "model" is effectively a % C-pointer (int32 or int64 within Matlab). Never change that variable, % or you may loose the LWPR model! %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
model = lwpr_storage('Store',model);
Step 2: Training
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Now we start training the model. We just call "lwpr_update" with one % input/output tuple after another. The function also returns the % current prediction for the input data, which we use to keep track % of the MSE on the training data % % For printing how the training proceeds, we also use % a) lwpr_num_data to report the number of training data the model % has seen % b) lwpr_num_rfs to report the number of receptive fields that % have been allocated so far %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
for k=1:20
ind = randperm(Ntr);
mse = 0;
for i=1:Ntr
[model,yp] = lwpr_update(model,Xtr(ind(i)),Ytr(ind(i)));
mse = mse + (Ytr(ind(i),:)-yp)^2;
end
nMSE = mse/Ntr/var(Ytr,1);
fprintf(1,'#Data: %5d #RFs: %3d nMSE=%5.3f\n', lwpr_num_data(model),
lwpr_num_rfs(model), nMSE);
end
A typical Matlab output for these few lines is printed below. In order to detect a poor choice of initial parameters, it is always useful to keep track of how many receptive fields are allocated, and how the normalised mean square error (nMSE) evolves:
#Data: 500 #RFs: 15 nMSE=0.301 #Data: 1000 #RFs: 19 nMSE=0.116 #Data: 1500 #RFs: 21 nMSE=0.041 #Data: 2000 #RFs: 21 nMSE=0.037 #Data: 2500 #RFs: 21 nMSE=0.034 #Data: 3000 #RFs: 21 nMSE=0.032 #Data: 3500 #RFs: 21 nMSE=0.031 #Data: 4000 #RFs: 21 nMSE=0.030 #Data: 4500 #RFs: 21 nMSE=0.029 #Data: 5000 #RFs: 21 nMSE=0.029 #Data: 5500 #RFs: 21 nMSE=0.029 #Data: 6000 #RFs: 21 nMSE=0.028 #Data: 6500 #RFs: 21 nMSE=0.028 #Data: 7000 #RFs: 21 nMSE=0.028 #Data: 7500 #RFs: 21 nMSE=0.028 #Data: 8000 #RFs: 21 nMSE=0.028 #Data: 8500 #RFs: 21 nMSE=0.027 #Data: 9000 #RFs: 21 nMSE=0.027 #Data: 9500 #RFs: 21 nMSE=0.027 #Data: 10000 #RFs: 21 nMSE=0.027
Step 3: Obtain predictions for unseen data
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % In the next lines, we retrieve predictions from the LWPR model % on the regularly sampled interval [0;10] % This is achieved by calling "lwpr_predict" % The first output is the prediction, the second is a % one-standard-deviation confidence bound. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Ntest = 500; Xtest = linspace(0,10,Ntest); Ytest = zeros(Ntest,1); Conf = zeros(Ntest,1); for k=1:500 [Ytest(k), Conf(k)] = lwpr_predict(model,Xtest(k)); end
% Plot the predictions and confidence bounds
plot(Xtest,Ytest,'b-'); hold on plot(Xtest,Ytest+Conf,'c-'); plot(Xtest,Ytest-Conf,'c-');
We sampled the input space regularly and evaluated the LWPR regression function and its confidence bounds. The resulting plot follows below. The prediction is indicated by a blue line, with a one-standard-deviation confidence interval depicted by the cyan lines.
Step 4a: Retrieve Matlab structure from MEX storage (cf. Step 1a)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % With the following line, we transform the MEX-internal LWPR model % back into a Matlab struct, and also free the internal memory. % After this, we can still use "model" to calculate predictions, % but this would be slower as before. However, we can now access all % internal variables, which we exploit to visualize the receptive % fields. % Note that this step is also necessary if you wish to save the % LWPR model together with the rest of you Matlab workspace, since % Matlab does not "know" about the internal storage scheme. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
model = lwpr_storage('GetFree',model);
Step 4b: Inspect LWPR model
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Any receptive field can be accessed by % model.sub(outDim).rfs(index) % where outDim is the output dimension (here only =1) and index % selects the RF within that output dimension. % Thus, model.sub(1).rfs(5).D corresponds to the distance metric % of the 5th receptive field. The field "c" denotes the center % of that field, beta0 is the offset, and beta the PLS regression % parameter (in 1-D equivalent to +/- the slope of the local model) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
t = linspace(0,2*pi,50); ct = cos(t); st = sin(t); for k=1:length(model.sub(1).rfs); r = 1/sqrt(model.sub(1).rfs(k).D); x = model.sub.rfs(k).c; y = model.sub.rfs(k).beta0; b = model.sub.rfs(k).beta * model.sub.rfs(k).U; plot(x+r*ct,r*st,'g-'); plot([x-r,x+r],[y-r*b,y+r*b],'k-'); end
The resulting plot looks like this:
Informatics Robot Lab Open House and Inaugural Lecture: Prof. Vijayakumar (Apr 25, 2012)
