Locally Weighted Projection Regression (LWPR) using rdyncall and the LWPR C library

lwpr-rdyncall.r

library(rdyncall)

dynbind("lwpr", "
        lwpr_init_model(*<LWPR_Model>iiZ)i;
        lwpr_duplicate_mode(*<LWPR_Model>*<LWPR_Model>)i;
        lwpr_set_init_alpha(*<LWPR_Model>d)i;
        lwpr_set_init_D(*<LWPR_Model>*dd)i;
        lwpr_set_init_D_diagonal(*<LWPR_Model>*d)i;
        lwpr_set_init_D_spherical(*<LWPR_Model>d)i;
        lwpr_update(*<LWPR_Model>*d*d*d*d)i;
        lwpr_predict(*<LWPR_Model>*dd*d*d*d)v;
        lwpr_write_xml(*<LWPR_Model>Z)i;
        lwpr_read_xml(*<LWPR_Model>Z*i)i;
        ")

parseStructInfos("LWPR_Model{iiii*d*d*ciidd*d*d*d*d*ddddddddiipp*d*d*d}nIn nInStore nOut n_data mean_x var_x name diag_only meta meta_rate penalty init_alpha norm_in norm_out init_D init_M w_gen w_prune init_lambda final_lambda tau_lambda init_S2 add_threshold kernel update_D sub ws storage xn yn")

testfunc = function(x) {
    10 * sin(7.8*log(1+x)) / (1 + 0.1*x**2)
}
 
Ntr = 500
Xtr = 10 * runif(Ntr)
Ytr = sapply(Xtr, testfunc) + 0.1 * rnorm(Ntr) * Xtr

model = new.struct(LWPR_Model)
lwpr_init_model(model, 1, 1, "Tutorial") 

lwpr_set_init_D(model, 20, 0)
model$update_D = 1
model$diag_only = TRUE
model$penalty = 0.0001
lwpr_set_init_alpha(model, 40)

for (i in 1:20) {
    idx = sample(Ntr,Ntr)
    for (j in idx) {
        yp=0
        lwpr_update(model, Xtr[j], Ytr[j], yp, NULL)
    }
}

Ntest = 500;
Xtest = seq(0,10,length.out=Ntest)
Ytest = numeric(Ntest)
Conf = numeric(Ntest) 
for (k in 1:Ntest) {
    yp = 0
    conf = 0
    lwpr_predict(model, Xtest[k], 0.0, yp, conf, NULL);
    Ytest[k] = yp
    Conf[k]  = conf
}

plot(Ytr ~ Xtr, pch=19, cex=0.5, col='red')
lines(Xtest , Ytest, col="blue") 
lines(Xtest , Ytest + Conf, col="cyan") 
lines(Xtest , Ytest - Conf, col="cyan")