warpfun: The 'warpfun' function

Description Usage Arguments Details Value Author(s) Examples

View source: R/warpfun.r

Description

This function is used to produce the warping function and learning the interpolation in the warped time domain.

Usage

1
warpfun(kkkrkg, bbb, peod, eps, fixlens, y_no, testData, witer)

Arguments

kkkrkg

'ode' class object.

bbb

list of 'rkhs' class object.

peod

vector(of length n_s) containing the period of warped signal. n_s is the length of the ode states.

eps

vector(of length n_s) containing the uncertainty level of the period. n_s is the length of the ode states.

fixlens

vector(of length n_s) containing the initial values of the hyper parameters of sigmoid basis function.

y_no

matrix(of size n_s*n_o) containing noisy observations. The row(of length n_s) represent the ode states and the column(of length n_o) represents the time points.

testData

vector(of size n_x) containing user defined time points which will be warped by the warping function.

witer

scale containing the number of iterations for optimising the hyper parameters of warping.

Details

Arguments of the 'warpfun' function are 'ode' class, 'rkhs' class, period of warped signal, uncertainty level of the period, initial values of the hyper parameters for sigmoid basis function, noisy observations and the time points that user want to warped. It return the interpolation for each of the ode states. The ode parameters are estimated using gradient matching, and the results are stored in the 'ode' class as the ode_par attribute.

Value

return list containing :

Author(s)

Mu Niu [email protected]

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
## Not run: 
require(mvtnorm)
noise = 0.1  
SEED = 19537
set.seed(SEED)
## Define ode function, we use lotka-volterra model in this example. 
## we have two ode states x[1], x[2] and four ode parameters alpha, beta, gamma and delta.
LV_fun = function(t,x,par_ode){
  alpha=par_ode[1]
  beta=par_ode[2]
  gamma=par_ode[3]
  delta=par_ode[4]
  as.matrix( c( alpha*x[1]-beta*x[2]*x[1] , -gamma*x[2]+delta*x[1]*x[2] ) )
}
## Define the gradient of ode function against ode parameters 
## df/dalpha,  df/dbeta, df/dgamma, df/ddelta where f is the differential equation.
LV_grlNODE= function(par,grad_ode,y_p,z_p) { 
alpha = par[1]; beta= par[2]; gamma = par[3]; delta = par[4]
dres= c(0)
dres[1] = sum( -2*( z_p[1,]-grad_ode[1,])*y_p[1,]*alpha ) 
dres[2] = sum( 2*( z_p[1,]-grad_ode[1,])*y_p[2,]*y_p[1,]*beta)
dres[3] = sum( 2*( z_p[2,]-grad_ode[2,])*gamma*y_p[2,] )
dres[4] = sum( -2*( z_p[2,]-grad_ode[2,])*y_p[2,]*y_p[1,]*delta)
dres
}

## create a ode class object
kkk0 = ode$new(2,fun=LV_fun,grfun=LV_grlNODE)
## set the initial values for each state at time zero.
xinit = as.matrix(c(0.5,1))
## set the time interval for the ode numerical solver.
tinterv = c(0,6)
## solve the ode numerically using predefined ode parameters. alpha=1, beta=1, gamma=4, delta=1.
kkk0$solve_ode(c(1,1,4,1),xinit,tinterv) 

## Add noise to the numerical solution of the ode model and use it as the noisy observation.
n_o = max( dim( kkk0$y_ode) )
t_no = kkk0$t
y_no =  t(kkk0$y_ode) + rmvnorm(n_o,c(0,0),noise*diag(2))

## create a ode class object by using the simulation data we created from the Ode numerical solver.
## If users have experiment data, they can replace the simulation data with the experiment data.
## set initial value of Ode parameters.
init_par = rep(c(0.1),4)
init_yode = t(y_no)
init_t = t_no
kkk = ode$new(1,fun=LV_fun,grfun=LV_grlNODE,t=init_t,ode_par= init_par, y_ode=init_yode )

## The following examples with CPU or elapsed time > 10s

## Use function 'rkg' to estimate the Ode parameters.
ktype ='rbf'
rkgres = rkg(kkk,y_no,ktype)
bbb = rkgres$bbb

###### warp all ode states
peod = c(6,5.3) ## the guessing period
eps= 1          ## the uncertainty level of period

###### learn the initial value of the hyper parameters of the warping basis function
fixlens=warpInitLen(peod,eps,rkgres)

kkkrkg = kkk$clone() ## make a copy of ode class objects
##learn the warping function, warp data points and do gradient matching in the warped time domain.
www = warpfun(kkkrkg,bbb,peod,eps,fixlens,y_no,kkkrkg$t)

dtilda= www$dtilda  ## gradient of warping function
bbbw = www$bbbw      ## interpolation in warped time domain
resmtest = www$wtime  ## warped time points
##display the results of parameter estimation using gradient matching in the warped time domain.
www$wkkk$ode_par       

## End(Not run)

KGode documentation built on March 18, 2018, 2:21 p.m.