tests/testthat/test-core-noise.R

In ngme2: Linear Latent Non-Gaussian Models with Flexible Distributions

test_that("test basic mn", {
  set.seed(100)
  n_obs <- 1000
  x <- rexp(n_obs)
  beta <- c(-1, 2)
  y <- beta[[1]] + x * beta[[2]] + rnorm(n_obs, sd = 0.5)

  # summary(lm(y~x))
  out <- ngme(
    y ~ x,
    data = data.frame(x = x, y = y),
    control_opt = control_opt(
      iterations = 100,
      # optimizer = precond_sgd(),
      print_check_info = FALSE,
      n_parallel_chain = 4
    )
  )
  expect_true(out$replicates[[1]]$noise$theta_sigma - log(1.5) < 1)
  out
  traceplot(out)

  # 2. nig case
  n_obs <- 1000
  x <- rnorm(n_obs)
  z <- rexp(n_obs)
  # beta <- c(-2, 3)
  mu <- 5
  sigma <- 0.8
  nu <- 0.1
  sim_noise <- rnig(n_obs, delta = -mu, mu = mu, nu = nu, sigma = sigma, seed = 10)
  mean(sim_noise)
  # y <- z*beta[[1]] + x * beta[[2]] + sim_noise - mean(sim_noise)
  y <- sim_noise - mean(sim_noise)

  out <- ngme(
    y ~ 0,
    data = data.frame(x = x, y = y), family = noise_nig(),
    control_opt = control_opt(
      iterations = 100,
      n_parallel_chain = 4,
      n_min_batch = 10,
      optimizer = precond_sgd(),
      print_check_info = FALSE
    )
  )
  out
  traceplot(out, hline = c(mu, sigma, nu, beta))
  expect_true(out$replicates[[1]]$noise$theta_sigma - log(2) < 1)
  expect_true(out$replicates[[1]]$noise$theta_mu - mu < 1)
  expect_true(out$replicates[[1]]$noise$theta_nu - log(nu) < 2)
  plot(noise_nig(mu = mu, nu = nu, sigma = sigma), out$replicates[[1]]$noise)
})

test_that("test ar1 (normal) + NIG noise", {
  set.seed(100)
  n_obs <- 1000
  rho <- 0.6
  sigma_ar <- 5
  model <- f(
    1:n_obs,
    model = ar1(rho = rho),
    noise = noise_normal(sigma = sigma_ar)
  )
  W <- simulate(model)[[1]]
  sd(W)

  x <- rexp(n_obs)
  mu <- 1
  sigma <- 1
  nu <- 0.5
  sim_noise <- rnig(n_obs, delta = -mu, mu = mu, nu = nu, sigma = sigma, seed = 10)
  sd(sim_noise)
  y <- W + sim_noise
  mean(sim_noise)

  out <- ngme(
    y ~ 0 + f(
      1:n_obs,
      model = ar1(),
      noise = noise_normal()
    ),
    data = data.frame(x = x, y = y),
    family = noise_nig(),
    control_opt = control_opt(
      iterations = 1000,
      n_parallel_chain = 4,
      # optimizer = precond_sgd(),
      n_min_batch = 10,
      # verbose=TRUE
    )
  )
  out
  traceplot(out, "field1", hline = c(rho, sigma_ar))
  traceplot(out, hline = c(mu, sigma, nu))

  expect_true(out$replicates[[1]]$noise$theta_sigma - log(sigma) < 2)
  expect_true(out$replicates[[1]]$noise$theta_mu - mu < 1)
  expect_true(out$replicates[[1]]$noise$theta_nu - log(nu) < 2)
})