demo/simulation-cfg1-demo.R

In frbl/OnlineSuperLearner: Online SuperLearner package

library('devtools')
library('magrittr')
library('doParallel')
library('foreach')
library('doParallel')
#load_all(".")

# Turn off asking for enter
par(ask=FALSE)

set.seed(12345)

## Make sure we use all cores
registerDoParallel(cores = parallel::detectCores() - 1)

log <- R.utils::Arguments$getVerbose(-1, timestamp=TRUE)

## Generate a dataset we will use for testing.
training_set_size <- 1e5
initial_data_size <-  500#training_set_size / 2
test_set_size <- 100

## What is the maximum number of iterations the OSL can use while going over the data?
## Note that in this case we split the data in equal parts with this number of iterations
max_iterations <- 20

## Specify the intervention we'd like to test, and also specify when we want to
## test this intervention
intervention <- list(variable = 'A', when = c(1), what = c(1))

tau <- 1

## B is the number of iterations we'll run before we hope to converge
B <- 100

# Initialize the simulator
#--------------------

## Generate observations for training
## These are used in the simulator / its scheme.
llW <- list(
  stochMech=function(numberOfBlocks) {
    rnorm(numberOfBlocks, 0, 0.1)
  },
  param=c(0, 0.5, -0.25, 0.1),
  rgen=identity
)

llA <- list(
  stochMech=function(ww) {
      rbinom(length(ww), 1, expit(ww))
  },
  param=c(-0.1, 0.1, 0.25),
  rgen=function(xx, delta=0.05){
    probability <- delta+(1-2*delta)*expit(xx)
    rbinom(length(xx), 1, probability)
  }
)

llY <- list(rgen={function(AW){
  aa <- AW[, "A"]
  ww <- AW[, grep("[^A]", colnames(AW))]
  mu <- aa*(0.4-0.2*sin(ww)+0.05*ww) +
    (1-aa)*(0.2+0.1*cos(ww)-0.03*ww)
  #mu <- aa * 0.9 + (1-aa) * 0.3
  rnorm(length(mu), mu, sd=0.01)}}
)

## Create a new simulator
sim <- Simulator.GAD$new(qw=llW, ga=llA, Qy=llY)

## Generate some fake data for testing and training
data.train <- sim$simulateWAY(training_set_size, verbose=log)
data.test <- sim$simulateWAY(test_set_size, verbose=log)

# Create the relevant variables
#------------------------------

## We'd like to use the following features in our estimation:
W <- RelevantVariable$new(formula = W ~ Y_lag_1 + A_lag_1 +  W_lag_1 + Y_lag_2, family = 'gaussian')
A <- RelevantVariable$new(formula = A ~ W + Y_lag_1 + A_lag_1 + W_lag_1, family = 'binomial')
Y <- RelevantVariable$new(formula = Y ~ A + W + Y_lag_1 + A_lag_1 +  W_lag_1 + Y_lag_2, family = 'gaussian')
relevantVariables <- c(W, A, Y)


## Define a list of algorithms to use
algos <- list()
#algos <- append(algos, list(list(algorithm = "ML.XGBoost",
                                 #params = list(nbins = c(5, 10, 15), online = TRUE))))

#algos <- append(algos, list(list(algorithm = "ML.NeuralNet",
                                 #params = list(nbins = c(5), online = TRUE))))

algos <- append(algos, list(list(algorithm = "ML.SpeedGLMSGD",
                                 params = list(nbins = c(50), online = TRUE))))
algos <- append(algos, list(list(algorithm = "ML.SpeedGLMSGD",
                                 params = list(nbins = c(15), online = TRUE))))
algos <- append(algos, list(list(algorithm = "ML.SpeedGLMSGD",
                                 params = list(nbins = c(5), online = TRUE))))
algos <- append(algos, list(list(algorithm = "ML.SpeedGLMSGD",
                                 algorithm_params = list(alpha = seq(0,1,0.2)),
                                 params = list(nbins = c(5), online = TRUE))))

## Fit the actual OSL
#--------------------
#unloadNamespace('OnlineSuperLearner'); unloadNamespace('condensier') ; install('../../osofr/condensier/', dependency=T);devtools::load_all('.')

#devtools::load_all('.');
osl <- OnlineSuperLearner::fit.OnlineSuperLearner(
  formulae = relevantVariables, ## Specify which are the formulae we expet
  data = data.train, ## Specify the data to train on
  algorithms = algos, ## SPecify the correct algorithms
  verbose = log, ## Logging information
  bounds = TRUE, ## Let the OSL generate the bounds based on the data it gets
  test_set_size = 5 + (3 * 3 + 3), ## The size of the minibatch test size. Note that for this test set size it is super important that at least enough observations are available as 
  initial_data_size =initial_data_size, ## Train the first iteration (Nl) on this part of the data
  max_iterations = max_iterations, ## Use at most max_iterations over the data
  mini_batch_size = (training_set_size / 2) / max_iterations, ## Split the remaining data into N-Nl/max_iterations equal blocks of data
  parallel = T
)

## Create a quick overview of the training curve (the risk over time)
OutputPlotGenerator.create_training_curve(
  osl$get_historical_cv_risk,
  relevantVariables = relevantVariables,
  output = 'curve1'
)

## First we simulate data given the intervention. That is, we specify in our
## simulation that we want to sample data when this intervention would be
## applied. After that we take the mean at tau, and as such approximate our
## treatment effect in the true population.
cat('Approximating truth...\n')
result.approx <- foreach(i=seq(B)) %do% {
  cat('Approximating truth in iteration (under intervention): ', i, '\n')
  data.int <- sim$simulateWAY(tau, qw = llW, ga = llA, Qy = llY,
                                  intervention = intervention, verbose = FALSE)
  data.int$Y[tau]
} %>% unlist
result.approx %>% mean

## The next step is to actually calculate the same intervention using the
## superlearner. We use a similar technique for this, as we try to calculate
## the mean of the intervention effects.

## We need to have data that includes the summary measures for the evaluation
## generate them here
data.train <- Data.Static$new(dataset = data.train)
osl$get_summary_measure_generator$set_trajectories(data.train)
data.train.set <- osl$get_summary_measure_generator$getNext(2)


## First we create the calculator to determine the intervention effects with.
intervention_effect_caluculator = InterventionEffectCalculator$new(
  bootstrap_iterations = B, 
  outcome_variable = Y$getY,
  verbose = log,
  parallel = TRUE
)

result <- lapply(c(TRUE, FALSE), function(discrete) {
  ## Actually evaluate the intervention for the discrete superlearner
  intervention_effect <- intervention_effect_caluculator$evaluate_single_intervention(
    osl = osl,
    intervention = intervention, 
    discrete = discrete, 
    initial_data = data.train.set$traj_1[1,],
    tau = tau
  ) %>% unlist
})

data <- list(truth = result.approx, dosl = result[[1]], osl = result[[2]])
OutputPlotGenerator.create_convergence_plot(data = data, output = 'convergence1')

cat('The effects of the interventions were:')
cat(paste('approx',':', result.approx %>% mean)) 
cat(paste('discre',':', result[[1]] %>% mean)) 
cat(paste('contin',':', result[[2]] %>% mean)) 

## Finally we run our kolmogorov smirnov test example to check whether we
## actually do a good job estimating the true conditional distributions.

## Define kolmogorov-smirnov test
T_iter <- 10
B_iter <- 1000
nbins <- 5

## Define the object that will be used to run the evalutation, and run the actual evaluations.
#devtools::load_all('.')
subject <- ConditionalDensityEvaluator$new(log, osl = osl, summary_measure_generator = osl$get_summary_measure_generator, cfg = 1)
result <- subject$evaluate(
  sim,
  T_iter, 
  B_iter,
  nbins = nbins,
  outcome_variable = Y
)

## Output the evaluation.
perc_significant <- subject$calculate_significance(result, TRUE, 0.05) %>% round(.,2)
paste(perc_significant,'% significant in the KS-test')