Gauss Newton method R - r

Find the MLE of the non-linear distribution (in R, using a Gauss-Newton method):
y = sin(x*theta) + epsilon
where epsilon ~ N(0 , 0.01^2)
To do this, I've been asked to generate some data that is uniformly (and randomly) distributed from 0 <= x <= 10 , with n = 200 and theta = 2 (just for generation).
For instance, values that are close to the maximum of the sin function (1, 4 etc.) will converge but others won't.
EDITED
I now understand what theta.iter means but I cannot seem to understand why it converges only sometimes and even then, which values to input to get a useful output of. Can someone explain?
theta <- 2
x <- runif(200, 0, 10)
x <- sort(x) #this is just to sort the generated data so that plotting it
#actually looks like a sine funciton
y <- sin(x*theta) + rnorm(200, mean = 0, sd = 0.1^2)
GN_sin <- function(theta.iter, x , y, epsilon){
index <- TRUE
while (index){
y.iter <- matrix(y - sin(x*theta.iter), 200, 1)
x.iter <- matrix(theta.iter*cos(x*theta.iter), 200, 1)
theta.new <- theta.iter +
solve(t(x.iter)%*%x.iter)%*%t(x.iter)%*%y.iter
if (abs(theta.new-theta.iter) < epsilon) {index <- FALSE}
theta.iter <- as.vector(theta.new)
cat(theta.iter, '\n')
}
}

Related

Generate a binary variable with a predefined correlation to an already existing variable

For a simulation study, I want to generate a set of random variables (both continuous and binary) that have predefined associations to an already existing binary variable, denoted here as x.
For this post, assume that x is generated following the code below. But remember: in real life, x is an already existing variable.
set.seed(1245)
x <- rbinom(1000, 1, 0.6)
I want to generate both a binary variable and a continuous variable. I have figured out how to generate a continuous variable (see code below)
set.seed(1245)
cor <- 0.8 #Correlation
y <- rnorm(1000, cor*x, sqrt(1-cor^2))
But I can't find a way to generate a binary variable that is correlated to the already existing variable x. I found several R packages, such as copula which can generate random variables with a given dependency structure. However, they do not provide a possibility to generate variables with a set dependency on an already existing variable.
Does anyone know how to do this in an efficient way?
Thanks!
If we look at the formula for correlation:
For the new vector y, if we preserve the mean, the problem is easier to solve. That means we copy the vector x and try to flip a equal number of 1s and 0s to achieve the intended correlation value.
If we let E(X) = E(Y) = x_bar , and E(XY) = xy_bar, then for a given rho, we simplify the above to:
(xy_bar - x_bar^2) / (x_bar - x_bar^2) = rho
Solve and we get:
xy_bar = rho * x_bar + (1-rho)*x_bar^2
And we can derive a function to flip a number of 1s and 0s to get the result:
create_vector = function(x,rho){
n = length(x)
x_bar = mean(x)
xy_bar = rho * x_bar + (1-rho)*x_bar^2
toflip = sum(x == 1) - round(n * xy_bar)
y = x
y[sample(which(x==0),toflip)] = 1
y[sample(which(x==1),toflip)] = 0
return(y)
}
For your example it works:
set.seed(1245)
x <- rbinom(1000, 1, 0.6)
cor(x,create_vector(x,0.8))
[1] 0.7986037
There are some extreme combinations of intended rho and p where you might run into problems, for example:
set.seed(111)
res = lapply(1:1000,function(i){
this_rho = runif(1)
this_p = runif(1)
x = rbinom(1000,1,this_p)
data.frame(
intended_rho = this_rho,
p = this_p,
resulting_cor = cor(x,create_vector(x,this_rho))
)
})
res = do.call(rbind,res)
ggplot(res,aes(x=intended_rho,y=resulting_cor,col=p)) + geom_point()
Here's a binomial one - the formula for q only depends on the mean of x and the correlation you desire.
set.seed(1245)
cor <- 0.8
x <- rbinom(100000, 1, 0.6)
p <- mean(x)
q <- 1/((1-p)/cor^2+p)
y <- rbinom(100000, 1, q)
z <- x*y
cor(x,z)
#> [1] 0.7984781
This is not the only way to do this - note that mean(z) is always less than mean(x) in this construction.
The continuous variable is even less well defined - do you really not care about its mean/variance, or anything else about its distibution?
Here's another simple version where it flips the variable both ways:
set.seed(1245)
cor <- 0.8
x <- rbinom(100000, 1, 0.6)
p <- mean(x)
q <- (1+cor/sqrt(1-(2*p-1)^2*(1-cor^2)))/2
y <- rbinom(100000, 1, q)
z <- x*y+(1-x)*(1-y)
cor(x,z)
#> [1] 0.8001219
mean(z)
#> [1] 0.57908

Machine Learning: Stochastic gradient descent for logistic regression in R: Calculating Eout and average number of epochs

I am trying to write a code to solve the following problem (As stated in HW5 in the CalTech course Learning from Data):
In this problem you will create your own target function f
(probability in this case) and data set D to see how Logistic
Regression works. For simplicity, we will take f to be a 0=1
probability so y is a deterministic function of x. Take d = 2 so you
can visualize the problem, and let X = [-1; 1]×[-1; 1] with uniform
probability of picking each x 2 X . Choose a line in the plane as the
boundary between f(x) = 1 (where y has to be +1) and f(x) = 0 (where y
has to be -1) by taking two random, uniformly distributed points from
X and taking the line passing through them as the boundary between y =
±1. Pick N = 100 training points at random from X , and evaluate the
outputs yn for each of these points xn. Run Logistic Regression with
Stochastic Gradient Descent to find g, and estimate Eout(the cross
entropy error) by generating a sufficiently large, separate set of
points to evaluate the error. Repeat the experiment for 100 runs with
different targets and take the average. Initialize the weight vector
of Logistic Regression to all zeros in each run. Stop the algorithm
when |w(t-1) - w(t)| < 0:01, where w(t) denotes the weight vector at
the end of epoch t. An epoch is a full pass through the N data points
(use a random permutation of 1; 2; · · · ; N to present the data
points to the algorithm within each epoch, and use different
permutations for different epochs). Use a learning rate of 0.01.
I am required to calculate the nearest value to Eout for N=100, and the average number of epochs for the required criterion.
I wrote and ran the code but I'm not getting the right answers (as stated in the solutions, these are Eout is near 0.1 and the number of epochs is near 350). The required number of epochs for a delta w of 0.01 comes to far too small (around 10), leaving the error too big (around 2). I then tried to replace the criterion with |w(t-1) - w(t)| < 0.001 (rather than 0.01). Then, the average required number of epochs was about 250 and out of sample error was about 0.35.
Is there something wrong with my code/solution, or is it possible that the answers provided are faulty? I've added comments to indicate what I intend to do at each step. Thanks in advance.
library(pracma)
h<- 0 # h will later be updated to number of required epochs
p<- 0 # p will later be updated to Eout
C <- matrix(ncol=10000, nrow=2) # Testing set, used to calculate out of sample error
d <- matrix(ncol=10000, nrow=1)
for(i in 1:10000){
C[, i] <- c(runif(2, min = -1, max = 1)) # Sample data
d[1, i] <- sign(C[2, i] - f(C[1, i]))
}
for(g in 1:100){ # 100 runs of the experiment
x <- runif(2, min = -1, max = 1)
y <- runif(2, min = -1, max = 1)
fit = (lm(y~x))
t <- summary(fit)$coefficients[,1]
f <- function(x){ # Target function
t[2]*x + t[1]
}
A <- matrix(ncol=100, nrow=2) # Sample data
b <- matrix(ncol=100, nrow=1)
norm_vec <- function(x) {sqrt(sum(x^2))} # vector norm calculator
w <- c(0,0) # weights initialized to zero
for(i in 1:100){
A[, i] <- c(runif(2, min = -1, max = 1)) # Sample data
b[1, i] <- sign(A[2, i] - f(A[1, i]))
}
q <- matrix(nrow = 2, ncol = 1000) # q tracks the weight vector at the end of each epoch
l= 1
while(l < 1001){
E <- function(z){ # cross entropy error function
x = z[1]
y = z[2]
v = z[3]
return(log(1 + exp(-v*t(w)%*%c(x, y))))
}
err <- function(xn1, xn2, yn){ #gradient of error function
return(c(-yn*xn1, -yn*xn2)*(exp(-yn*t(w)*c(xn1,xn2))/(1+exp(-yn*t(w)*c(xn1,xn2)))))
}
e = matrix(nrow = 2, ncol = 100) # e will track the required gradient at each data point
e[,1:100] = 0
perm = sample(100, 100, replace = FALSE, prob = NULL) # Random permutation of the data indices
for(j in 1:100){ # One complete Epoch
r = A[,perm[j]] # pick the perm[j]th entry in A
s = b[perm[j]] # pick the perm[j]th entry in b
e[,perm[j]] = err(r[1], r[2], s) # Gradient of the error
w = w - 0.01*e[,perm[j]] # update the weight vector accorng to the formula involving step size, gradient
}
q[,l] = w # the lth entry is the weight vector at the end of the lth epoch
if(l > 1 & norm_vec(q[,l] - q[,l-1])<0.001){ # given criterion to terminate the algorithm
break
}
l = l+1 # move to the next epoch
}
for(n in 1:10000){
p[g] = mean(E(c(C[1,n], C[2, n], d[n]))) # average over 10000 data points, of the error function, in experiment no. g
}
h[g] = l #gth entry in the vector h, tracks the number of epochs in the gth iteration of the experiment
}
mean(h) # Mean number of epochs needed
mean(p) # average Eout, over 100 experiments

Simulated dataset in R

I'm simulating another dataset here, and am stuck again!
Here's what I want to do:
200 observations, with 90 independent variables (mean 0, sd 1)
the equation to create y is: y = 2x_1 + ... + 2x_30 - x_31 - ... - x_60 + 0*x_61 + ... + 0*x_90 + mu
(In other words, the first 30 x values will have a coefficient of 2, next 30 values have coefficient of -1 and last 30 values have coefficient of 0). mu is also a random generated normal variable with mean 0, sd 10.
Here's what I have so far:
set.seed(11)
n <- 200
mu <- rnorm(200,0,10)
p1 <- for(i in 1:200){
rnorm(200,0,1)
}
p2 <- cbind(p1)
p3 <- for(i in 1:90){
if i<=30, y=2x
if i>30 & i<=60, y=-x
if i>60 & i<=90, y=0x
}
I'm still learning many aspects of R, so I'm pretty sure the code has much wrong with it, even in terms of syntax. Your help would really be appreciated!
Thanks!
Try
library(mvtnorm)
coefs <- rep(c(2, -1, 0), each=30)
mu <- rnorm(200, 0, 10)
m <- rep(0, 90) # mean of independent variables
sig <- diag(90) # cov of indep variables
x <- rmvnorm(200, mean=m, sigma=sig) # generates 200 observations from multivariate normal
y <- x%*%coefs + mu
In case, if you are not comfortable with linear-algebra
n <- 200
coefs <- rep(c(2, -1, 0), each=30)
mu <- rnorm(n, 0, 10)
x <- matrix(nrow=n, ncol=90) # initializes the indep.vars
for(i in 1:90){
x[, i] <- rnorm(200, 0, 1)
}
y <- rep(NA, n) # initializes the dependent vars
for(i in 1:n){
y[i] = sum(x[i,]*coefs) + m[i]
}
x[i,]*coefs gives exactly (2*x_1,..., 2*x_30, -x_31,...,- x_60,0*x_61,...,0*x_90) because * is element-wise operation.
You'd better learn the rudimentaries of R, before actually doing something with it.

Adding two random variables via convolution in R

I would like to compute the convolution of two probability distributions in R and I need some help. For the sake of simplicity, let's say I have a variable x that is normally distributed with mean = 1.0 and stdev = 0.5, and y that is log-normally distributed with mean = 1.5 and stdev = 0.75. I want to determine z = x + y. I understand that the distribution of z is not known a priori.
As an aside the real world example I am working with requires addition to two random variables that are distributed according to a number of different distributions.
Does anyone know how to add two random variables by convoluting the probability density functions of x and y?
I have tried generating n normally distributed random values (with above parameters) and adding them to n log-normally distributed random values. However, I wish to know if I can use the convolution method instead. Any help would be greatly appreciated.
EDIT
Thank you for these answers. I define a pdf, and try to do the convolution integral, but R complains on the integration step. My pdfs are Log Pearson 3 and are as follows
dlp3 <- function(x, a, b, g) {
p1 <- 1/(x*abs(b) * gamma(a))
p2 <- ((log(x)-g)/b)^(a-1)
p3 <- exp(-1* (log(x)-g) / b)
d <- p1 * p2 * p3
return(d)
}
f.m <- function(x) dlp3(x,3.2594,-0.18218,0.53441)
f.s <- function(x) dlp3(x,9.5645,-0.07676,1.184)
f.t <- function(z) integrate(function(x,z) f.s(z-x)*f.m(x),-Inf,Inf,z)$value
f.t <- Vectorize(f.t)
integrate(f.t, lower = 0, upper = 3.6)
R complains at the last step since the f.t function is bounded and my integration limits are probably not correct. Any ideas on how to solve this?
Here is one way.
f.X <- function(x) dnorm(x,1,0.5) # normal (mu=1.5, sigma=0.5)
f.Y <- function(y) dlnorm(y,1.5, 0.75) # log-normal (mu=1.5, sigma=0.75)
# convolution integral
f.Z <- function(z) integrate(function(x,z) f.Y(z-x)*f.X(x),-Inf,Inf,z)$value
f.Z <- Vectorize(f.Z) # need to vectorize the resulting fn.
set.seed(1) # for reproducible example
X <- rnorm(1000,1,0.5)
Y <- rlnorm(1000,1.5,0.75)
Z <- X + Y
# compare the methods
hist(Z,freq=F,breaks=50, xlim=c(0,30))
z <- seq(0,50,0.01)
lines(z,f.Z(z),lty=2,col="red")
Same thing using package distr.
library(distr)
N <- Norm(mean=1, sd=0.5) # N is signature for normal dist
L <- Lnorm(meanlog=1.5,sdlog=0.75) # same for log-normal
conv <- convpow(L+N,1) # object of class AbscontDistribution
f.Z <- d(conv) # distribution function
hist(Z,freq=F,breaks=50, xlim=c(0,30))
z <- seq(0,50,0.01)
lines(z,f.Z(z),lty=2,col="red")
I was having trouble getting integrate() to work for different density parameters, so I came up with an alternative to #jlhoward's using Riemann approximation:
set.seed(1)
#densities to be convolved. could also put these in the function below
d1 <- function(x) dnorm(x,1,0.5) #
d2 <- function(y) dlnorm(y,1.5, 0.75)
#Riemann approximation of convolution
conv <- function(t, a, b, d) { #a to b needs to cover the range of densities above. d needs to be small for accurate approx.
z <- NA
x <- seq(a, b, d)
for (i in 1:length(t)){
print(i)
z[i] <- sum(d1(x)*d2(t[i]-x)*d)
}
return(z)
}
#check against sampled convolution
X <- rnorm(1000, 1, 0.5)
Y <- rlnorm(1000, 1.5, 0.75)
Z <- X + Y
t <- seq(0, 50, 0.05) #range to evaluate t, smaller increment -> smoother curve
hist(Z, breaks = 50, freq = F, xlim = c(0,30))
lines(t, conv(t, -100, 100, 0.1), type = "s", col = "red")

Trouble estimating E[u(X)] by simulation when u() is exponential

I'm trying to use R to estimate E[u(X)] where u is a utility function and X is a random variable. More specifically, I want to be able to rank E[u(X)] and E[u(Y)] for two random variables X and Y -- only the ranking matters.
My problem is that u(x) = -exp(-sigma * x) for some sigma > 0, and this converges very rapidly to zero. So I have many cases where I expect, say, E[u(X)] > E[u(Y)], but because they are so close to zero, my simulation cannot distinguish them.
Does anyone have any advice for me?
I am only interested in ranking the two expected utilities, so u(x) can be replaced by any u.tilde(x) = a * u(x) + b, where a > 0 and b can be any number.
Below is an example where X and Y are both normal (in which case I think there is a closed form solution, but pretend X and Y have complicated distributions that I can only simulate from).
get.u <- function(sigma=1) {
stopifnot(sigma > 0)
utility <- function(x) {
return(-exp(-sigma * x))
}
return(utility)
}
u <- get.u(sigma=1)
curve(u, from=0, to=10) # Converges very rapidly to zero
n <- 10^4
x <- rnorm(n, 10^4, sd=10)
y <- rnorm(n, 10^4, sd=10^3)
mean(u(x)) == mean(u(y)) # Returns True (they're both 0), but I expect E[u(x)] > E[u(y)]
## An example of replacing u with a*u + b
get.scaled.u <- function(sigma=1) {
stopifnot(sigma > 0) # Risk averse
utility <- function(x) {
return(-exp(-sigma * x + sigma * 10^4))
}
return(utility)
}
u <- get.scaled.u(sigma=1)
mean(u(x)) > mean(u(y)) # True as desired
x <- rnorm(n, 10^4, sd=10^3)
y <- rnorm(n, 10^4, sd=2*10^3)
mean(u(x)) > mean(u(y)) # False again -- they're both -Inf
Is finding a clever way to scale u the correct way to deal with this problem? For example, suppose X and Y both have bounded support -- if I know the bounds, how can I scale u to guarantee that a*u + b will be neither too close to -Inf, nor too close to zero?
Edit: I didn't know about multiple precision packages. Rmpfr is helpful:
library(Rmpfr)
x.precise <- mpfr(x, 100)
y.precise <- mpfr(y, 100)
mean(u(x.precise)) > mean(u(y.precise)) # True

Resources