Given two point clouds I want to find for each point from the first point cloud the nearest neighbour from the second point cloud. Also, each pair of neighbours should be unique. The solution was already given here for Python. However, I was wondering if a similar approach exists for R (I'd like to avoid the suggested cvxpy library from the Python solution which requires the pyscipopt library which again requires the installation of the SCIP Optimization Suite).
Some example code with two point clouds:
set.seed(666)
# Example data
px = runif(210, min = 0, max = 100)
py = runif(210, min = 0, max = 100)
pc1 = cbind(x = px[1:100], y = py[1:100])
pc2 = cbind(x = px[101:210], y = py[101:210])
plot(pc1, pch = 16, col = 1)
points(pc2, pch = 16, col = 2)
# Calculate distance matrix
# library(pdist)
# d = pdist(pc1, pc2)
# d = as.matrix(d)
# Find closest neighbour
library(FNN)
nn = get.knnx(pc2, pc1, k = 1)
for(i in 1:nrow(pc1)) lines(x = c(pc1[i,1], pc2[nn$nn.index[i,1],1]),
y = c(pc1[i,2], pc2[nn$nn.index[i,1],2]))
As shown in the image above, I'm capable of finding the nearest neighbour for each point from Point Cloud 1 (black) to Point Cloud 2 (red). Yet, multiple points from Point Cloud 2 are assigned to the same point from Point Cloud 1. Any idea how to find unique pairs with minimal overall distances instead?
Edit:
I tried another approach by finding the closest pair iteratively and removing that pair from the following queries:
# Approach 2:
pairs = matrix(NA, ncol = 4, nrow = nrow(pc1)) #storage for pairs
colnames(pairs) = c("x(pc1)", "y(pc1)", "x(pc2)", "y(pc2)")
pc2_copy = pc2 # copy of Point Cloud 2 which will shrink each iteration
for(i in 1:nrow(pc1)){
nn = get.knnx(pc2_copy, pc1[i,,drop = FALSE], k = 1)
pairs[i,1:2] = pc1[i,1:2,drop = FALSE]
pairs[i,3:4] = pc2_copy[nn$nn.index[1,1],1:2,drop = FALSE]
pc2_copy = pc2_copy[-c(nn$nn.index[1,1]),] #remove the corresponding point from the matrix
}
plot(pc1, pch = 16, col = 1)
points(pc2, pch = 16, col = 2)
for(i in 1:nrow(pairs)) lines(x = pairs[i,c(1,3)], y = pairs[i, c(2,4)])
While this gives me unique pairs, I don't believe this is anywhere near an ideal solution (in my real data example some distances are very little while others are immense with an obviously much better solution by eye).
The package RcppHungarian will solve this type of assignment problem using the Hungarian algorithm:
set.seed(666)
# Example data
px = runif(210, min = 0, max = 100)
py = runif(210, min = 0, max = 100)
pc1 = cbind(x = px[1:100], y = py[1:100])
pc2 = cbind(x = px[101:210], y = py[101:210])
nn <- RcppHungarian::HungarianSolver(
proxy::dist( # distance matrix
pc1, pc2, method = "euclidean"
)
)$pairs
any(duplicated(nn[,2]))
#> [1] FALSE
plot(pc1, pch = 16, col = 1)
points(pc2, pch = 16, col = 2)
for(i in 1:nrow(pc1)) lines(x = c(pc1[i,1], pc2[nn[i,2],1]),
y = c(pc1[i,2], pc2[nn[i,2],2]))
I am trying to fit a variety of (truncated) probability distributions to the same very thin set of quantiles. I can do it but it seems to require lots of duplication of the same code. Is there a neater way?
I am using this code by Nadarajah and Kotz to generate the pdf of the truncated distributions:
qtrunc <- function(p, spec, a = -Inf, b = Inf, ...)
{
tt <- p
G <- get(paste("p", spec, sep = ""), mode = "function")
Gin <- get(paste("q", spec, sep = ""), mode = "function")
tt <- Gin(G(a, ...) + p*(G(b, ...) - G(a, ...)), ...)
return(tt)
}
where spec can be the name of any untruncated distribution for which code in R exists, and the ... argument is used to provide the names of the parameters of that untruncated distribution.
To achieve the best fit I need to measure the distance between the given quantiles and those calculated using arbitrary values of the parameters of the distribution. In the case of the gamma distribution, for example, the code is as follows:
spec <- "gamma"
fit_gamma <- function(x, l = 0, h = 20, t1 = 5, t2 = 13){
ct1 <- qtrunc(p = 1/3, spec, a = l, b = h, shape = x[1],rate = x[2])
ct2 <- qtrunc(p = 2/3, spec, a = l, b = h, shape = x[1],rate = x[2])
dist <- vector(mode = "numeric", length = 2)
dist[1] <- (t1 - ct1)^2
dist[2] <- (t2- ct2)^2
return(sqrt(sum(dist)))
}
where l is the lower truncation, h is the higher and I am given the two tertiles t1 and t2.
Finally, I seek the best fit using optim, thus:
gamma_fit <- optim(par = c(2, 4),
fn = fit_gamma,
l = l,
h = h,
t1 = t1,
t2 = t2,
method = "L-BFGS-B",
lower = c(1.01, 1.4)
Now suppose I want to do the same thing but fitting a normal distribution instead. The names of the parameters of the normal distribution that I am using in R are mean and sd.
I can achieve what I want but only by writing a whole new function fit_normal that is extremely similar to my fit_gamma function but with the new parameter names used in the definition of ct1 and ct2.
The problem of duplication of code becomes very severe because I wish to try fitting a large number of different distributions to my data.
What I want to know is whether there is a way of writing a generic fit_spec as it were so that the parameter names do not have to be written out by me.
Use x as a named list to create a list of arguments to pass into qtrunc() using do.call().
fit_distro <- function(x, spec, l = 0, h = 20, t1 = 5, t2 = 13){
args <- c(x, list(spec = spec, a = l, b = h))
ct1 <- do.call(qtrunc, args = c(list(p = 1/3), args))
ct2 <- do.call(qtrunc, args = c(list(p = 2/3), args))
dist <- vector(mode = "numeric", length = 2)
dist[1] <- (t1 - ct1)^2
dist[2] <- (t2 - ct2)^2
return(sqrt(sum(dist)))
}
This is called as follows, which is the same as your original function.
fit_distro(list(shape = 2, rate = 3), "gamma")
# [1] 13.07425
fit_gamma(c(2, 3))
# [1] 13.07425
This will work with other distributions, for however many parameters they have.
fit_distro(list(mean = 10, sd = 3), "norm")
# [1] 4.08379
fit_distro(list(shape1 = 2, shape2 = 3, ncp = 10), "beta")
# [1] 12.98371
We simulated a data set and created a model.
set.seed(459)
# seed mass
n <- 1000
seed.mass <- round(rnorm(n, mean = 250, sd = 75),digits = 1)
## Setting up the deterministic function
detFunc <- function(a,b,x){
return(exp(a+b*x)) / (1+exp(a+b*x))
}
# logit link function for the binomial
inv.link <- function(z){
p <-1/(1+exp(-z))
return(p)
}
#setting a and b values
a <- -2.109
b <- 0.02
# Simulating data
germination <- (rbinom(n = n, size = 10,
p = inv.link(detFunc(x = seed.mass, a = a, b = b))
))/10
## make data frame
mydata <- data.frame("predictor" = seed.mass, "response" = germination)
# plotting the data
tmp.x <- seq(0,1e3,length.out=500)
plot(germination ~ seed.mass,
xlab = "seed mass (mg)",
ylab = "germination proportion")
lines(tmp.x,inv.link(detFunc(x = tmp.x, a = a, b = b)),col="red",lwd=2)
When we check the model we created and infer the parameters, we get an error:
Error in optim(par = c(a = -2.109, b = 0.02), fn = function (p) : initial value in 'vmmin' is not finite
library(bbmle)
mod1<-mle2(response ~ dbinom(size = 10,
p = inv.link(detFunc(x = predictor, a = a, b = b))
),
data = mydata,
start = list("a"= -2.109 ,"b"= 0.02))
We're stumped and can't figure out why we're getting this error.
Your problem is that you're trying to fit a binomial outcome (which must be an integer) to a proportion.
You can use round(response*10) as your predictor (to put the proportion back on the count scale; round() is because (a/b)*b is not always exactly equal to a in floating-point math ...) Specifically, with your setup
mod1 <- mle2(round(response*10) ~ dbinom(size = 10,
p = inv.link(detFunc(x = predictor, a = a, b = b))
),
data = mydata,
start = list(a = -2.109 ,b = 0.02))
works fine. coef(mod1) is {-1.85, 0.018}, plausibly close to the true values you started with (we don't expect to recover the true values exactly, except as the average of many simulations [and even then MLE is only asymptotically unbiased, i.e. for large data sets ...]
The proximal problem is that trying to evaluate dbinom() with a non-integer value gives NA. The full output from your model fit would have been:
Error in optim(par = c(a = -2.109, b = 0.02), fn = function (p) :
initial value in 'vmmin' is not finite
In addition: There were 50 or more warnings (use warnings() to see the first 50)
It's always a good idea to check those additional warnings ... in this case they are all of the form
1: In dbinom(x = c(1, 1, 1, 0.8, 1, 1, 1, 1, 1, 1, 1, 0.8, ... :
non-integer x = 0.800000
which might have given you a clue ...
PS you can use qlogis() and plogis() from base R for your link and inverse-link functions ...
I am working on replicating the scoring rule found in a paper Forecasting the intermittent demand for slow-moving inventories: A modelling approach
The paper describes the scoring rule as follows:
This is my attempt
y <- rpois(n = 100, lambda = 10) # forecasted distribution
x <- 10 # actual value
drps_score <- function(x = value, y = q){
# x = actual value (single observation); y = quantile forecasted value (vector)
Fy = ecdf(y) # cdf function
indicator <- ifelse(y - x > 0, 1, 0) # Heaviside
score <- sum((indicator - Fy(y))^2)
return(score)
}
> drps_score(x = x, y = y)
[1] 53.028
This seems to work well until I provide a vector of 0s as follows:
y <- rep(x = 0, 100)
> drps_score(x = x, y = y)
[1] 0
I know that one of their methods used in this paper was a 0s forecast and their results did not show 0 for DRPS. This makes me think that the calculation is off.
I think there are a few issues at play here.
First off, I don't think you are computing the correct sum inside the scoring function. The score asks you to sum across all possible values of y (i.e. across all positive integers) not across all forecasted samples of y.
Second, I don't think the above definition gives the desired result, with \hat F (y) defined to be 0 when y=x then you don't get a zero score for a forecast with a point mass at the true value. (Yes, I'm saying that source is "wrong", or at least has a definition that doesn't give the desired result.) Here is a re-formulated function that I think fixes both issues:
x <- 10 # actual value
drps_score <- function(x = value, y = q, nsum=100){
# x = actual value (single observation); y = quantile forecasted value (vector)
Fy = ecdf(y) # cdf function
ysum <- 0:nsum
indicator <- ifelse(ysum - x >= 0, 1, 0) # Heaviside
score <- sum((indicator - Fy(ysum))^2)
return(score)
}
> drps_score(x = x, y = rpois(n = 1000, lambda = 8))
[1] 1.248676
> drps_score(x = x, y = rpois(n = 1000, lambda = 9))
[1] 0.878183
> drps_score(x = x, y = rpois(n = 1000, lambda = 10))
[1] 0.692667
> drps_score(x = x, y = rep(10, 100))
[1] 0
> drps_score(x = x, y = rpois(n = 1000, lambda = 11))
[1] 0.883333
The above shows that the distribution that is centered on the true value (lambda=10) has the lowest score for distributions that aren't a point mass.
I'm using Sutton & Barto's ebook Reinforcement Learning: An Introduction to study reinforcement learning. I'm having some issues trying to emulate the results (plots) on the action-value page.
More specifically, how can I simulate the greedy value for each task? The book says:
...we can plot the performance and behavior of various methods as
they improve with experience over 1000 plays...
So I guess I have to keep track of the exploratory values as better ones are found. The issue is how to do this using the greedy approach - since there are no exploratory moves, how do I know what is a greedy behavior?
Thanks for all the comments and answers!
UPDATE: See code on my answer.
I finally got this right. The eps player should beat the greedy player because of the exploratory moves, as pointed out int the book.
The code is slow and need some optimizations, but here it is:
get.testbed = function(arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1){
optimal = rnorm(arms, u, sdev.arm)
rewards = sapply(optimal, function(x)rnorm(plays, x, sdev.rewards))
list(optimal = optimal, rewards = rewards)
}
play.slots = function(arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1, eps = 0.1){
testbed = get.testbed(arms, plays, u, sdev.arm, sdev.rewards)
optimal = testbed$optimal
rewards = testbed$rewards
optim.index = which.max(optimal)
slot.rewards = rep(0, arms)
reward.hist = rep(0, plays)
optimal.hist = rep(0, plays)
pulls = rep(0, arms)
probs = runif(plays)
# vetorizar
for (i in 1:plays){
## dont use ifelse() in this case
## idx = ifelse(probs[i] < eps, sample(arms, 1), which.max(slot.rewards))
idx = if (probs[i] < eps) sample(arms, 1) else which.max(slot.rewards)
reward.hist[i] = rewards[i, idx]
if (idx == optim.index)
optimal.hist[i] = 1
slot.rewards[idx] = slot.rewards[idx] + (rewards[i, idx] - slot.rewards[idx])/(pulls[idx] + 1)
pulls[idx] = pulls[idx] + 1
}
list(slot.rewards = slot.rewards, reward.hist = reward.hist, optimal.hist = optimal.hist, pulls = pulls)
}
do.simulation = function(N = 100, arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1, eps = c(0.0, 0.01, 0.1)){
n.players = length(eps)
col.names = paste('eps', eps)
rewards.hist = matrix(0, nrow = plays, ncol = n.players)
optim.hist = matrix(0, nrow = plays, ncol = n.players)
colnames(rewards.hist) = col.names
colnames(optim.hist) = col.names
for (p in 1:n.players){
for (i in 1:N){
play.results = play.slots(arms, plays, u, sdev.arm, sdev.rewards, eps[p])
rewards.hist[, p] = rewards.hist[, p] + play.results$reward.hist
optim.hist[, p] = optim.hist[, p] + play.results$optimal.hist
}
}
rewards.hist = rewards.hist/N
optim.hist = optim.hist/N
optim.hist = apply(optim.hist, 2, function(x)cumsum(x)/(1:plays))
### Plot helper ###
plot.result = function(x, n.series, colors, leg.names, ...){
for (i in 1:n.series){
if (i == 1)
plot.ts(x[, i], ylim = 2*range(x), col = colors[i], ...)
else
lines(x[, i], col = colors[i], ...)
grid(col = 'lightgray')
}
legend('topleft', leg.names, col = colors, lwd = 2, cex = 0.6, box.lwd = NA)
}
### Plot helper ###
#### Plots ####
require(RColorBrewer)
colors = brewer.pal(n.players + 3, 'Set2')
op <-par(mfrow = c(2, 1), no.readonly = TRUE)
plot.result(rewards.hist, n.players, colors, col.names, xlab = 'Plays', ylab = 'Average reward', lwd = 2)
plot.result(optim.hist, n.players, colors, col.names, xlab = 'Plays', ylab = 'Optimal move %', lwd = 2)
#### Plots ####
par(op)
}
To run it just call
do.simulation(N = 100, arms = 10, eps = c(0, 0.01, 0.1))
You could also choose to make use of the R package "contextual", which aims to ease the implementation and evaluation of both context-free (as described in Sutton & Barto) and contextual (such as for example LinUCB) Multi-Armed Bandit policies.
The package actually offers a vignette on how to replicate all Sutton & Barto bandit plots. For example, to generate the ε-greedy plots, just simulate EpsilonGreedy policies against a Gaussian bandit :
library(contextual)
set.seed(2)
mus <- rnorm(10, 0, 1)
sigmas <- rep(1, 10)
bandit <- BasicGaussianBandit$new(mu_per_arm = mus, sigma_per_arm = sigmas)
agents <- list(Agent$new(EpsilonGreedyPolicy$new(0), bandit, "e = 0, greedy"),
Agent$new(EpsilonGreedyPolicy$new(0.1), bandit, "e = 0.1"),
Agent$new(EpsilonGreedyPolicy$new(0.01), bandit, "e = 0.01"))
simulator <- Simulator$new(agents = agents, horizon = 1000, simulations = 2000)
history <- simulator$run()
plot(history, type = "average", regret = FALSE, lwd = 1, legend_position = "bottomright")
plot(history, type = "optimal", lwd = 1, legend_position = "bottomright")
Full disclosure: I am one of the developers of the package.
this is what I have so far based on our chat:
set.seed(1)
getRewardsGaussian <- function(arms, plays) {
## assuming each action has a normal distribution
# first generate new means
QStar <- rnorm(arms, 0, 1)
# then for each mean, generate `play`-many samples
sapply(QStar, function(u)
rnorm(plays, u, 1))
}
CalculateRewardsPerMethod <- function(arms=7, epsi1=0.01, epsi2=0.1
, plays=1000, methods=c("greedy", "epsi1", "epsi2")) {
# names for easy handling
names(methods) <- methods
arm.names <- paste0("Arm", ifelse((1:arms)<10, 0, ""), 1:arms)
# this could be different if not all actions' rewards have a gaussian dist.
rewards.source <- getRewardsGaussian(arms, plays)
# Three dimensional array to track running averages of each method
running.avgs <-
array(0, dim=c(plays, arms, length(methods))
, dimnames=list(PlayNo.=NULL, Arm=arm.names, Method=methods))
# Three dimensional array to track the outcome of each play, according to each method
rewards.received <-
array(NA_real_, dim=c(plays, 2, length(methods))
, dimnames=list(PlayNo.=seq(plays), Outcome=c("Arm", "Reward"), Method=methods))
# define the function internally to not have to pass running.avgs
chooseAnArm <- function(p) {
# Note that in a tie, which.max returns the lowest value, which is what we want
maxes <- apply(running.avgs[p, ,methods, drop=FALSE], 3, which.max)
# Note: deliberately drawing two separate random numbers and keeping this as
# two lines of code to accent that the two draws should not be related
if(runif(1) < epsi1)
maxes["epsi1"] <- sample(arms, 1)
if(runif(1) < epsi2)
maxes["epsi2"] <- sample(arms, 1)
return(maxes)
}
## TODO: Perform each action at least once, then select according to algorithm
## Starting points. Everyone starts at machine 3
choice <- c(3, 3, 3)
reward <- rewards.source[1, choice]
## First run, slightly different
rewards.received[1,,] <- rbind(choice, reward)
running.avgs[1, choice, ] <- reward # if different starting points, this needs to change like below
## HERE IS WHERE WE START PULLING THE LEVERS ##
## ----------------------------------------- ##
for (p in 2:plays) {
choice <- chooseAnArm(p)
reward <- rewards.source[p, choice]
# Note: When dropping a dim, the methods will be the columns
# and the Outcome info will be the rows. Use `rbind` instead of `cbind`.
rewards.received[p,,names(choice)] <- rbind(choice, reward)
## Update the running averages.
## For each method, the current running averages are the same as the
## previous for all arms, except for the one chosen this round.
## Thus start with last round's averages, then update the one arm.
running.avgs[p,,] <- running.avgs[p-1,,]
# The updating is only involved part (due to lots of array-indexing)
running.avgs[p,,][cbind(choice, 1:3)] <-
sapply(names(choice), function(m)
# Update the running average for the selected arm (for the current play & method)
mean( rewards.received[ 1:p,,,drop=FALSE][ rewards.received[1:p,"Arm",m] == choice[m],"Reward",m])
)
} # end for-loop
## DIFFERENT RETURN OPTIONS ##
## ------------------------ ##
## All rewards received, in simplifed matrix (dropping information on arm chosen)
# return(rewards.received[, "Reward", ])
## All rewards received, along with which arm chosen:
# return(rewards.received)
## Running averages of the rewards received by method
return( apply(rewards.received[, "Reward", ], 2, cumsum) / (1:plays) )
}
### EXECUTION (AND SIMULATION)
## PARAMETERS
arms <- 10
plays <- 1000
epsi1 <- 0.01
epsi2 <- 0.1
simuls <- 50 # 2000
methods=c("greedy", "epsi1", "epsi2")
## Single Iteration:
### we can run system time to get an idea for how long one will take
tme <- system.time( CalculateRewardsPerMethod(arms=arms, epsi1=epsi1, epsi2=epsi2, plays=plays) )
cat("Expected run time is approx: ", round((simuls * tme[["elapsed"]]) / 60, 1), " minutes")
## Multiple iterations (simulations)
rewards.received.list <- replicate(simuls, CalculateRewardsPerMethod(arms=arms, epsi1=epsi1, epsi2=epsi2, plays=plays), simplify="array")
## Compute average across simulations
rewards.received <- apply(rewards.received.list, 1:2, mean)
## RESULTS
head(rewards.received, 17)
MeanRewards <- rewards.received
## If using an alternate return method in `Calculate..` use the two lines below to calculate running avg
# CumulRewards <- apply(rewards.received, 2, cumsum)
# MeanRewards <- CumulRewards / (1:plays)
## PLOT
plot.ts(MeanRewards[, "greedy"], col = 'red', lwd = 2, ylim = range(MeanRewards), ylab = 'Average reward', xlab="Plays")
lines(MeanRewards[, "epsi1"], col = 'orange', lwd = 2)
lines(MeanRewards[, "epsi2"], col = 'navy', lwd = 2)
grid(col = 'darkgray')
legend('bottomright', c('greedy', paste("epsi1 =", epsi1), paste("epsi2 =", epsi2)), col = c('red', 'orange', 'navy'), lwd = 2, cex = 0.8)
You may also want to check this link
https://www.datahubbs.com/multi_armed_bandits_reinforcement_learning_1/
Copy of the relevant code from the above source
It does not use R but simply np.random.rand() from numpy
class eps_bandit:
'''
epsilon-greedy k-bandit problem
Inputs
=====================================================
k: number of arms (int)
eps: probability of random action 0 < eps < 1 (float)
iters: number of steps (int)
mu: set the average rewards for each of the k-arms.
Set to "random" for the rewards to be selected from
a normal distribution with mean = 0.
Set to "sequence" for the means to be ordered from
0 to k-1.
Pass a list or array of length = k for user-defined
values.
'''
def __init__(self, k, eps, iters, mu='random'):
# Number of arms
self.k = k
# Search probability
self.eps = eps
# Number of iterations
self.iters = iters
# Step count
self.n = 0
# Step count for each arm
self.k_n = np.zeros(k)
# Total mean reward
self.mean_reward = 0
self.reward = np.zeros(iters)
# Mean reward for each arm
self.k_reward = np.zeros(k)
if type(mu) == list or type(mu).__module__ == np.__name__:
# User-defined averages
self.mu = np.array(mu)
elif mu == 'random':
# Draw means from probability distribution
self.mu = np.random.normal(0, 1, k)
elif mu == 'sequence':
# Increase the mean for each arm by one
self.mu = np.linspace(0, k-1, k)
def pull(self):
# Generate random number
p = np.random.rand()
if self.eps == 0 and self.n == 0:
a = np.random.choice(self.k)
elif p < self.eps:
# Randomly select an action
a = np.random.choice(self.k)
else:
# Take greedy action
a = np.argmax(self.k_reward)
reward = np.random.normal(self.mu[a], 1)
# Update counts
self.n += 1
self.k_n[a] += 1
# Update total
self.mean_reward = self.mean_reward + (
reward - self.mean_reward) / self.n
# Update results for a_k
self.k_reward[a] = self.k_reward[a] + (
reward - self.k_reward[a]) / self.k_n[a]
def run(self):
for i in range(self.iters):
self.pull()
self.reward[i] = self.mean_reward
def reset(self):
# Resets results while keeping settings
self.n = 0
self.k_n = np.zeros(k)
self.mean_reward = 0
self.reward = np.zeros(iters)
self.k_reward = np.zeros(k)