context("GPModel_gaussian_process") # Avoid that long tests get executed on CRAN if(Sys.getenv("GPBOOST_ALL_TESTS") == "GPBOOST_ALL_TESTS"){ TOLERANCE_LOOSE <- 1E-2 TOLERANCE_STRICT <- 1E-6 DEFAULT_OPTIM_PARAMS <- list(optimizer_cov = "gradient_descent", lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, optimizer_coef = "gradient_descent", lr_coef = 0.1, convergence_criterion = "relative_change_in_log_likelihood") DEFAULT_OPTIM_PARAMS_STD <- c(DEFAULT_OPTIM_PARAMS, list(std_dev = TRUE)) # Function that simulates uniform random variables sim_rand_unif <- function(n, init_c=0.1){ mod_lcg <- 2^32 # modulus for linear congruential generator (random0 used) sim <- rep(NA, n) sim[1] <- floor(init_c * mod_lcg) for(i in 2:n) sim[i] <- (22695477 * sim[i-1] + 1) %% mod_lcg return(sim / mod_lcg) } # Create data n <- 100 # number of samples # Simulate locations / features of GP d <- 2 # dimension of GP locations coords <- matrix(sim_rand_unif(n=n*d, init_c=0.1), ncol=d) D <- as.matrix(dist(coords)) # Simulate GP sigma2_1 <- 1^2 # marginal variance of GP rho <- 0.1 # range parameter Sigma <- sigma2_1 * exp(-D/rho) + diag(1E-20,n) C <- t(chol(Sigma)) b_1 <- qnorm(sim_rand_unif(n=n, init_c=0.8)) eps <- as.vector(C %*% b_1) # Random coefficients Z_SVC <- matrix(sim_rand_unif(n=n*2, init_c=0.6), ncol=2) # covariate data for random coeffients colnames(Z_SVC) <- c("var1","var2") b_2 <- qnorm(sim_rand_unif(n=n, init_c=0.17)) b_3 <- qnorm(sim_rand_unif(n=n, init_c=0.42)) eps_svc <- as.vector(C %*% b_1 + Z_SVC[,1] * C %*% b_2 + Z_SVC[,2] * C %*% b_3) # Error term xi <- qnorm(sim_rand_unif(n=n, init_c=0.1)) / 5 # Data for linear mixed effects model X <- cbind(rep(1,n),sin((1:n-n/2)^2*2*pi/n)) # design matrix / covariate data for fixed effect beta <- c(2,2) # regression coefficients # cluster_ids cluster_ids <- c(rep(1,0.4*n),rep(2,0.6*n)) # GP with multiple observations at the same locations coords_multiple <- matrix(sim_rand_unif(n=n*d/4, init_c=0.1), ncol=d) coords_multiple <- rbind(coords_multiple,coords_multiple,coords_multiple,coords_multiple) D_multiple <- as.matrix(dist(coords_multiple)) Sigma_multiple <- sigma2_1*exp(-D_multiple/rho)+diag(1E-10,n) C_multiple <- t(chol(Sigma_multiple)) b_multiple <- qnorm(sim_rand_unif(n=n, init_c=0.8)) eps_multiple <- as.vector(C_multiple %*% b_multiple) test_that("Gaussian process model ", { y <- eps + xi # Estimation using gradient descent and Nesterov acceleration gp_model <- GPModel(gp_coords = coords, cov_function = "exponential") capture.output( fit(gp_model, y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') cov_pars <- c(0.03784221, 0.07943467, 1.07390943, 0.25351519, 0.11451432, 0.03840236) num_it <- 59 expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_equal(dim(gp_model$get_cov_pars())[2], 3) expect_equal(dim(gp_model$get_cov_pars())[1], 2) expect_equal(gp_model$get_num_optim_iter(), num_it) # Can switch between likelihoods gp_model <- GPModel(gp_coords = coords, cov_function = "exponential") gp_model$set_likelihood("gamma") gp_model$set_likelihood("gaussian") capture.output( fit(gp_model, y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) # Gradient descent without Nesterov acceleration params <- DEFAULT_OPTIM_PARAMS_STD params$use_nesterov_acc <- FALSE capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = params) , file='NUL') cov_pars_other <- c(0.04040441, 0.08036674, 1.06926607, 0.25360131, 0.11502362, 0.03877014) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_other)),5E-6) expect_equal(gp_model$get_num_optim_iter(), 97) # Using a too large learning rate params <- DEFAULT_OPTIM_PARAMS_STD params$lr_cov <- 1 capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = params) , file='NUL') cov_pars_other <- c(0.04487369, 0.08285696, 1.07537253, 0.25676579, 0.11566763, 0.03928107) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_other)), TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 27) # Different terminations criterion params <- DEFAULT_OPTIM_PARAMS_STD params$convergence_criterion = "relative_change_in_parameters" capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = params), file='NUL') cov_pars_other_crit <- c(0.03276547, 0.07715343, 1.07617676, 0.25177603, 0.11352557, 0.03770062) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_other_crit)), TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 382) ll <- gp_model$neg_log_likelihood(y=y,cov_pars=gp_model$get_cov_pars()[1,]) expect_lt(abs(ll-122.7752664),TOLERANCE_STRICT) # Fisher scoring params <- DEFAULT_OPTIM_PARAMS_STD params$optimizer_cov = "fisher_scoring" params$lr_cov <- 1 capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = params), file='NUL') cov_pars_fisher <- c(0.03300593, 0.07725225, 1.07584118, 0.25180563, 0.11357012, 0.03773325) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_fisher)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 8) # Nelder-mead capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "nelder_mead", delta_rel_conv=1e-6)) , file='NUL') cov_pars_est <- as.vector(gp_model$get_cov_pars()) expect_lt(sum(abs(cov_pars_est-cov_pars[c(1,3,5)])),TOLERANCE_LOOSE) expect_equal(gp_model$get_num_optim_iter(), 40) # Test default values for delta_rel_conv for nelder_mead capture.output( gp_model_default <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "nelder_mead")) , file='NUL') capture.output( gp_model_8 <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "nelder_mead", delta_rel_conv=1e-8)) , file='NUL') expect_false(isTRUE(all.equal(gp_model_default$get_cov_pars(), gp_model$get_cov_pars()))) expect_true(isTRUE(all.equal(gp_model_default$get_cov_pars(), gp_model_8$get_cov_pars()))) # Test default values for delta_rel_conv for gradient_descent capture.output( gp_model_default <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "gradient_descent")) , file='NUL') capture.output( gp_model_8 <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "gradient_descent", delta_rel_conv=1e-8)) , file='NUL') capture.output( gp_model_6 <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "gradient_descent", delta_rel_conv=1e-6)) , file='NUL') expect_true(isTRUE(all.equal(gp_model_default$get_cov_pars(), gp_model_6$get_cov_pars()))) expect_false(isTRUE(all.equal(gp_model_default$get_cov_pars(), gp_model_8$get_cov_pars()))) # BFGS capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "bfgs")), file='NUL') cov_pars_est <- as.vector(gp_model$get_cov_pars()) expect_lt(sum(abs(cov_pars_est-cov_pars[c(1,3,5)])),TOLERANCE_LOOSE) expect_equal(gp_model$get_num_optim_iter(), 16) # Adam capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "adam")), file='NUL') cov_pars_est <- as.vector(gp_model$get_cov_pars()) expect_lt(sum(abs(cov_pars_est-cov_pars[c(1,3,5)])),TOLERANCE_LOOSE) expect_equal(gp_model$get_num_optim_iter(), 498) # Prediction from fitted model capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = list(optimizer_cov = "fisher_scoring", delta_rel_conv = 1E-6, use_nesterov_acc = FALSE, convergence_criterion = "relative_change_in_parameters")), file='NUL') coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) expect_error(predict(gp_model))# coord data not provided pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expected_mu <- c(0.06960478, 1.61299381, 0.44053480) expected_cov <- c(6.218737e-01, 2.024102e-05, 2.278875e-07, 2.024102e-05, 3.535390e-01, 8.479210e-07, 2.278875e-07, 8.479210e-07, 4.202154e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Prediction of variances only pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) # Predict training data random effects training_data_random_effects <- predict_training_data_random_effects(gp_model, predict_var = TRUE) preds <- predict(gp_model, gp_coords_pred = coords, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(training_data_random_effects[,1] - preds$mu)),1E-6) expect_lt(sum(abs(training_data_random_effects[,2] - preds$var)),1E-6) # Prediction using given parameters gp_model <- GPModel(gp_coords = coords, cov_function = "exponential") cov_pars_pred = c(0.02,1.2,0.9) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE) expected_mu <- c(0.08704577, 1.63875604, 0.48513581) expected_cov <- c(1.189093e-01, 1.171632e-05, -4.172444e-07, 1.171632e-05, 7.427727e-02, 1.492859e-06, -4.172444e-07, 1.492859e-06, 8.107455e-02) cov_no_nugget <- expected_cov cov_no_nugget[c(1,5,9)] <- expected_cov[c(1,5,9)] - cov_pars_pred[1] expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Prediction of variances only pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-cov_no_nugget)),1E-6) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.1,1.6,0.2),y=y) expect_lt(abs(nll-124.2549533),1E-6) # Do optimization using optim and e.g. Nelder-Mead gp_model <- GPModel(gp_coords = coords, cov_function = "exponential") opt <- optim(par=c(0.1,2,0.2), fn=gp_model$neg_log_likelihood, y=y, method="Nelder-Mead") expect_lt(sum(abs(opt$par-cov_pars[c(1,3,5)])),TOLERANCE_LOOSE) expect_lt(abs(opt$value-(122.7752694)),1E-5) expect_equal(as.integer(opt$counts[1]), 198) # Other covariance functions capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 0.5, y = y, params = DEFAULT_OPTIM_PARAMS_STD) , file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), num_it) capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 1.5, y = y, params = DEFAULT_OPTIM_PARAMS_STD) , file='NUL') cov_pars_other <- c(0.22926543, 0.08486055, 0.87886348, 0.24059253, 0.10726402, 0.01542898) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_other)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 16) capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 2.5, y = y, params = DEFAULT_OPTIM_PARAMS_STD) , file='NUL') cov_pars_other <- c(0.27251105, 0.08316755, 0.83205621, 0.23561744, 0.10536460, 0.01062167) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_other)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 13) }) test_that("Gaussian process model with linear regression term ", { y <- eps + X%*%beta + xi # Fit model gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, X=X, params = list(optimizer_cov = "fisher_scoring", optimizer_coef = "wls", delta_rel_conv = 1E-6, use_nesterov_acc = FALSE, std_dev = TRUE, convergence_criterion = "relative_change_in_parameters")) cov_pars <- c(0.008461342, 0.069973492, 1.001562822, 0.214358560, 0.094656409, 0.029400407) coef <- c(2.30780026, 0.21365770, 1.89951426, 0.09484768) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_lt(sum(abs( as.vector(gp_model$get_coef())-coef)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 11) # Prediction coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) X_test <- cbind(rep(1,3),c(-0.5,0.2,0.4)) pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE) expected_mu <- c(1.196952, 4.063324, 3.156427) expected_cov <- c(6.305383e-01, 1.358861e-05, 8.317903e-08, 1.358861e-05, 3.469270e-01, 2.686334e-07, 8.317903e-08, 2.686334e-07, 4.255400e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Gradient descent gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y=y, X=X, params = list(optimizer_cov = "gradient_descent", optimizer_coef = "gradient_descent", delta_rel_conv = 1E-6, use_nesterov_acc = TRUE, std_dev = FALSE, lr_coef=1)) cov_pars <- c(0.01624576, 0.99717015, 0.09616822) coef <- c(2.305484, 1.899207) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 99) # Nelder-Mead gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, X=X, params = list(optimizer_cov = "nelder_mead", optimizer_coef = "nelder_mead", maxit=1000, delta_rel_conv = 1e-12)) cov_pars <- c(0.008459373, 1.001564796, 0.094655964) coef <- c(2.307798, 1.899516) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 429) # BFGS gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, X=X, params = list(optimizer_cov = "bfgs", maxit=1000)) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-2) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),1E-2) expect_gt(gp_model$get_num_optim_iter(), 24)# different compilers result in slightly different results expect_lt(gp_model$get_num_optim_iter(), 27) # Adam gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, X=X, params = list(optimizer_cov = "adam", maxit=5000)) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-2) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),1E-2) expect_gt(gp_model$get_num_optim_iter(), 1979) # different compilers result in slightly different results expect_lt(gp_model$get_num_optim_iter(), 2008) }) test_that("Gaussian process and two random coefficients ", { y <- eps_svc + xi # Fit model capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_rand_coef_data = Z_SVC, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, maxit=10)), file='NUL') expected_values <- c(0.25740068, 0.22608704, 0.83503539, 0.41896403, 0.15039055, 0.10090869, 1.61010233, 0.84207763, 0.09015444, 0.07106099, 0.25064640, 0.62279880, 0.08720822, 0.32047865) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-expected_values)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 10) # Predict training data random effects cov_pars <- gp_model$get_cov_pars()[1,] training_data_random_effects <- predict_training_data_random_effects(gp_model) Z_SVC_test <- cbind(rep(0,length(y)),rep(0,length(y))) preds <- predict(gp_model, gp_coords_pred = coords, gp_rand_coef_data_pred=Z_SVC_test, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(training_data_random_effects[,1] - preds$mu)),1E-6) Z_SVC_test <- cbind(rep(1,length(y)),rep(0,length(y))) preds2 <- predict(gp_model, gp_coords_pred = coords, gp_rand_coef_data_pred=Z_SVC_test, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(training_data_random_effects[,2] - (preds2$mu - preds$mu))),1E-6) Z_SVC_test <- cbind(rep(0,length(y)),rep(1,length(y))) preds3 <- predict(gp_model, gp_coords_pred = coords, gp_rand_coef_data_pred=Z_SVC_test, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(training_data_random_effects[,3] - (preds3$mu - preds$mu))),1E-6) # Prediction gp_model <- GPModel(gp_coords = coords, gp_rand_coef_data = Z_SVC, cov_function = "exponential") coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) Z_SVC_test <- cbind(c(0.1,0.3,0.7),c(0.5,0.2,0.4)) expect_error(gp_model$predict(y = y, gp_coords_pred = coord_test, cov_pars = c(0.1,1,0.1,0.8,0.15,1.1,0.08)))# random slope data not provided pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, gp_rand_coef_data_pred=Z_SVC_test, cov_pars = c(0.1,1,0.1,0.8,0.15,1.1,0.08), predict_cov_mat = TRUE) expected_mu <- c(-0.1669209, 1.6166381, 0.2861320) expected_cov <- c(9.643323e-01, 3.536846e-04, -1.783557e-04, 3.536846e-04, 5.155009e-01, 4.554321e-07, -1.783557e-04, 4.554321e-07, 7.701614e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Predict variances pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, gp_rand_coef_data_pred=Z_SVC_test, cov_pars = c(0.1,1,0.1,0.8,0.15,1.1,0.08), predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_rand_coef_data = Z_SVC, y = y, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc= FALSE, maxit=5)), file='NUL') expected_values <- c(0.000242813, 0.176573955, 1.008181385, 0.397341267, 0.141084495, 0.070671768, 1.432715033, 0.708039197, 0.055598038, 0.048988825, 0.430573036, 0.550644708, 0.038976112, 0.106110593) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-expected_values)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 5) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.1,1,0.1,0.8,0.15,1.1,0.08),y=y) expect_lt(abs(nll-149.4422184),1E-5) }) test_that("Gaussian process model with cluster_id's not constant ", { y <- eps + xi capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, cluster_ids = cluster_ids, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.05414149, 0.08722111, 1.05789166, 0.22886740, 0.12702368, 0.04076914) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 247) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, cluster_ids = cluster_ids, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.05414149, 0.08722111, 1.05789166, 0.22886740, 0.12702368, 0.04076914) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-5) expect_equal(gp_model$get_num_optim_iter(), 20) # Prediction coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) cluster_ids_pred = c(1,3,1) gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", cluster_ids = cluster_ids) pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cluster_ids_pred = cluster_ids_pred, cov_pars = c(0.1,1,0.15), predict_cov_mat = TRUE) expected_mu <- c(-0.01437506, 0.00000000, 0.93112902) expected_cov <- c(0.743055189, 0.000000000, -0.000140644, 0.000000000, 1.100000000, 0.000000000, -0.000140644, 0.000000000, 0.565243468) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Predict variances pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cluster_ids_pred = cluster_ids_pred, cov_pars = c(0.1,1,0.15), predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) }) test_that("Gaussian process model with multiple observations at the same location ", { y <- eps_multiple + xi capture.output( gp_model <- fitGPModel(gp_coords = coords_multiple, cov_function = "exponential", y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, maxit = 500)), file='NUL') cov_pars <- c(0.037145465, 0.006065652, 1.151982610, 0.434770575, 0.191648634, 0.102375515) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 12) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords_multiple, cov_function = "exponential", y = y, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.037136462, 0.006064181, 1.153630335, 0.435788570, 0.192080613, 0.102631006) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-5) expect_equal(gp_model$get_num_optim_iter(), 14) # Predict training data random effects training_data_random_effects <- predict_training_data_random_effects(gp_model, predict_var = TRUE) preds <- predict(gp_model, gp_coords_pred = coords_multiple, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(training_data_random_effects[,1] - preds$mu)),1E-6) expect_lt(sum(abs(training_data_random_effects[,2] - preds$var)),1E-6) # Prediction coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) gp_model <- GPModel(gp_coords = coords_multiple, cov_function = "exponential") pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cov_pars = c(0.1,1,0.15), predict_cov_mat = TRUE) expected_mu <- c(-0.1460550, 1.0042814, 0.7840301) expected_cov <- c(0.6739502109, 0.0008824337, -0.0003815281, 0.0008824337, 0.6060039551, -0.0004157361, -0.0003815281, -0.0004157361, 0.7851787946) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Predict variances pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cov_pars = c(0.1,1,0.15), predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) }) test_that("Vecchia approximation for Gaussian process model ", { y <- eps + xi params_vecchia <- list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters") # Maximal number of neighbors capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, vecchia_ordering = "none"), file='NUL') capture.output( fit(gp_model, y = y, params = params_vecchia), file='NUL') cov_pars <- c(0.03276544, 0.07715339, 1.07617623, 0.25177590, 0.11352557, 0.03770062) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(dim(gp_model$get_cov_pars())[2], 3) expect_equal(dim(gp_model$get_cov_pars())[1], 2) expect_equal(gp_model$get_num_optim_iter(), 382) # Same thing without Vecchia approximation capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential"), file='NUL') capture.output( fit(gp_model, y = y, params = params_vecchia), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(dim(gp_model$get_cov_pars())[2], 3) expect_equal(dim(gp_model$get_cov_pars())[1], 2) expect_equal(gp_model$get_num_optim_iter(), 382) # Random ordering capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, vecchia_ordering="random", y = y, params = params_vecchia), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(dim(gp_model$get_cov_pars())[2], 3) expect_equal(dim(gp_model$get_cov_pars())[1], 2) expect_equal(gp_model$get_num_optim_iter(), 382) # Prediction using given parameters capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, vecchia_ordering = "none"), file='NUL') coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) cov_pars = c(0.02,1.2,0.9) gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all", num_neighbors_pred=n+2) pred <- predict(gp_model, y = y, gp_coords_pred = coord_test, cov_pars = cov_pars, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08704577, 1.63875604, 0.48513581) expected_cov <- c(1.189093e-01, 1.171632e-05, -4.172444e-07, 1.171632e-05, 7.427727e-02, 1.492859e-06, -4.172444e-07, 1.492859e-06, 8.107455e-02) exp_cov_no_nugget <- expected_cov exp_cov_no_nugget[c(1,5,9)] <- expected_cov[c(1,5,9)] - cov_pars[1] expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred <- predict(gp_model, y = y, gp_coords_pred = coord_test, cov_pars = cov_pars, predict_cov_mat = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-exp_cov_no_nugget)),1E-6) # Prediction of variances only pred <- predict(gp_model, y = y, gp_coords_pred = coord_test, cov_pars = cov_pars, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred <- predict(gp_model, y = y, gp_coords_pred = coord_test, cov_pars = cov_pars, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(as.vector(pred$var)-exp_cov_no_nugget[c(1,5,9)])),1E-6) # Vecchia approximation with 30 neighbors capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering = "none"), file='NUL') capture.output( fit(gp_model, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, maxit=1000, convergence_criterion = "relative_change_in_parameters")) , file='NUL') cov_pars_vecchia <- c(0.03297349, 0.07716447, 1.07691542, 0.25221730, 0.11378505, 0.03782172) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_vecchia)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 378) # Prediction from fitted model coord_test <- cbind(c(0.1,0.10001,0.7),c(0.9,0.90001,0.55)) gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_obs_only", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expected_mu_vecchia <- c(0.06968068, 0.06967750, 0.44208925) expected_cov_vecchia <- c(0.6214955, 0.0000000, 0.0000000, 0.0000000, 0.6215069, 0.0000000, 0.0000000, 0.0000000, 0.4199531) expect_lt(sum(abs(pred$mu-expected_mu_vecchia)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov_vecchia)),1E-6) # Vecchia approximation with 30 neighbors and random ordering capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering="random"), file='NUL') capture.output( fit(gp_model, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, maxit=1000, convergence_criterion = "relative_change_in_parameters")) , file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_vecchia)),TOLERANCE_LOOSE) expect_gt(gp_model$get_num_optim_iter(), 368) # different compilers result in slightly different results expect_lt(gp_model$get_num_optim_iter(), 379) # Prediction from fitted model gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_obs_only", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expect_lt(sum(abs(pred$mu-expected_mu_vecchia)),TOLERANCE_LOOSE) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov_vecchia)),TOLERANCE_LOOSE) # Predict training data random effects training_data_random_effects <- predict_training_data_random_effects(gp_model, predict_var = TRUE) gp_model$set_prediction_data(vecchia_pred_type = "latent_order_obs_first_cond_obs_only") preds <- predict(gp_model, gp_coords_pred = coords, predict_response = FALSE, predict_var = TRUE) expect_lt(sum(abs(training_data_random_effects[,1] - preds$mu)),1E-3) expect_lt(sum(abs(training_data_random_effects[,2] - preds$var)),1E-3) # Fisher scoring & default ordering capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering="none", y = y, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, maxit=100, convergence_criterion = "relative_change_in_parameters")), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_vecchia)),TOLERANCE_LOOSE) expect_equal(gp_model$get_num_optim_iter(), 16) # expect_gt(gp_model$get_num_optim_iter(), 16) # different compilers result in slightly different results # expect_lt(gp_model$get_num_optim_iter(), 21) # Prediction using given parameters cov_pars_pred <- c(0.02,1.2,0.9) gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_obs_only", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08665472, 0.08664854, 0.49011216) expected_cov <- c(0.11891, 0.00000000, 0.00000000, 0.00000000, 0.1189129, 0.00000000, 0.00000000, 0.00000000, 0.08108126) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred_var <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred_var$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred_var$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred_var2 <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred_var$mu - pred_var2$mu)),1E-6) expect_lt(sum(abs(pred_var$var - cov_pars_pred[1] - pred_var2$var)),1E-6) # Prediction with vecchia_pred_type = "order_obs_first_cond_all" gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08665472, 0.08661259, 0.49011216) expected_cov <- c(0.11891004, 0.09889262, 0.00000000, 0.09889262, 0.11891291, 0.00000000, 0.00000000, 0.00000000, 0.08108126) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred_var <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred_var$mu-expected_mu)),TOLERANCE_LOOSE) expect_lt(sum(abs(as.vector(pred_var$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred_var2 <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred_var$mu - pred_var2$mu)),1E-6) expect_lt(sum(abs(pred_var$var - cov_pars_pred[1] - pred_var2$var)),1E-6) # Prediction with vecchia_pred_type = "order_pred_first" gp_model$set_prediction_data(vecchia_pred_type = "order_pred_first", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08498682, 0.08502034, 0.49572748) expected_cov <- c(1.189037e-01, 9.888624e-02, -1.080005e-05, 9.888624e-02, 1.189065e-01, -1.079431e-05, -1.080005e-05, -1.079431e-05, 8.101757e-02) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred_var <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred_var$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred_var$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred_var2 <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred_var$mu - pred_var2$mu)),1E-6) expect_lt(sum(abs(pred_var$var - cov_pars_pred[1] - pred_var2$var)),1E-6) # Prediction with vecchia_pred_type = "latent_order_obs_first_cond_obs_only" gp_model$set_prediction_data(vecchia_pred_type = "latent_order_obs_first_cond_obs_only", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08616985, 0.08616384, 0.48721314) expected_cov <- c(1.189100e-01, 7.324225e-03, -5.851427e-07, 7.324225e-03, 1.189129e-01, -5.850749e-07, -5.851427e-07, -5.850750e-07, 8.107749e-02) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred_var <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred_var$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred_var$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred_var2 <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred_var$mu - pred_var2$mu)),1E-6) expect_lt(sum(abs(pred_var$var - cov_pars_pred[1] - pred_var2$var)),1E-6) # Prediction with vecchia_pred_type = "latent_order_obs_first_cond_all" gp_model$set_prediction_data(vecchia_pred_type = "latent_order_obs_first_cond_all", num_neighbors_pred = 30) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_cov_mat = TRUE, predict_response = TRUE) expected_mu <- c(0.08616985, 0.08616377, 0.48721314) expected_cov <- c(1.189100e-01, 9.889258e-02, -5.851418e-07, 9.889258e-02, 1.189129e-01, -5.850764e-07, -5.851418e-07, -5.850764e-07, 8.107749e-02) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) pred_var <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = TRUE) expect_lt(sum(abs(pred_var$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred_var$var)-expected_cov[c(1,5,9)])),1E-6) # Predict latent process pred_var2 <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = cov_pars_pred, predict_var = TRUE, predict_response = FALSE) expect_lt(sum(abs(pred_var$mu - pred_var2$mu)),1E-6) expect_lt(sum(abs(pred_var$var - cov_pars_pred[1] - pred_var2$var)),1E-6) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.1,1.6,0.2),y=y) expect_lt(abs(nll-124.2252524),1E-6) }) test_that("Vecchia approximation for Gaussian process model with linear regression term ", { y <- eps + X%*%beta + xi # Fit model capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n+2, vecchia_ordering = "none", y = y, X=X, params = list(optimizer_cov = "fisher_scoring", optimizer_coef = "wls", std_dev = TRUE, delta_rel_conv = 1E-6, use_nesterov_acc = FALSE, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.008461342, 0.069973492, 1.001562822, 0.214358560, 0.094656409, 0.029400407) coef <- c(2.30780026, 0.21365770, 1.89951426, 0.09484768) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 11) # Prediction coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) X_test <- cbind(rep(1,3),c(-0.5,0.2,0.4)) gp_model$set_prediction_data(vecchia_pred_type = "latent_order_obs_first_cond_all") capture.output( pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE, predict_response = TRUE) , file='NUL') expected_mu <- c(1.196952, 4.063324, 3.156427) expected_cov <- c(6.305383e-01, 1.358861e-05, 8.317903e-08, 1.358861e-05, 3.469270e-01, 2.686334e-07, 8.317903e-08, 2.686334e-07, 4.255400e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) }) test_that("Vecchia approximation for Gaussian process model with cluster_id's not constant ", { y <- eps + xi capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering = "none", y = y, cluster_ids = cluster_ids, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.05, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.05374602, 0.08709594, 1.05800024, 0.22867128, 0.12680152, 0.04066888) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 474) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering = "none", y = y, cluster_ids = cluster_ids, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-5) expect_equal(gp_model$get_num_optim_iter(), 20) # Prediction coord_test <- cbind(c(0.1,0.2,0.1001),c(0.9,0.4,0.9001)) cluster_ids_pred = c(1,3,1) capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering = "none", cluster_ids = cluster_ids), file='NUL') gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all", num_neighbors_pred = 30) capture.output( pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cluster_ids_pred = cluster_ids_pred, cov_pars = c(0.1,1,0.15), predict_cov_mat = TRUE), file='NUL') expected_mu <- c(-0.01438585, 0.00000000, -0.01500132) expected_cov <- c(0.7430552, 0.0000000, 0.6423148, 0.0000000, 1.1000000, 0.0000000, 0.6423148, 0.0000000, 0.7434589) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) }) test_that("Vecchia approximation for Gaussian process model with multiple observations at the same location ", { y <- eps_multiple + xi capture.output( gp_model <- fitGPModel(gp_coords = coords_multiple, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, y = y, vecchia_ordering = "none", params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, maxit = 500)), file='NUL') cov_pars <- c(0.037145465, 0.006065652, 1.151982610, 0.434770575, 0.191648634, 0.102375515) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-4) expect_equal(gp_model$get_num_optim_iter(), 12) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords_multiple, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, y = y, vecchia_ordering = "none", params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6, convergence_criterion = "relative_change_in_parameters")), file='NUL') cov_pars <- c(0.037136462, 0.006064181, 1.153630335, 0.435788570, 0.192080613, 0.102631006) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-5) expect_equal(gp_model$get_num_optim_iter(), 14) # Prediction coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) cluster_ids_pred = c(1,3,1) capture.output( gp_model <- GPModel(gp_coords = coords_multiple, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n+2, vecchia_ordering = "none"), file='NUL') gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all") capture.output( pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, cov_pars = c(0.1,1,0.15), predict_cov_mat = TRUE), file='NUL') expected_mu <- c(-0.1460550, 1.0042814, 0.7840301) expected_cov <- c(0.6739502109, 0.0008824337, -0.0003815281, 0.0008824337, 0.6060039551, -0.0004157361, -0.0003815281, -0.0004157361, 0.7851787946) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) }) test_that("Vecchia approximation for Gaussian process and two random coefficients ", { y <- eps_svc + xi # Fit model using gradient descent with Nesterov acceleration capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n-1, gp_rand_coef_data = Z_SVC, vecchia_ordering = "none", y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5, maxit=10)), file='NUL') expected_values <- c(0.25740068, 0.22608704, 0.83503539, 0.41896403, 0.15039055, 0.10090869, 1.61010233, 0.84207763, 0.09015444, 0.07106099, 0.25064640, 0.62279880, 0.08720822, 0.32047865) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-expected_values)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 10) # Prediction capture.output( gp_model <- GPModel(gp_coords = coords, gp_rand_coef_data = Z_SVC, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = n+2, vecchia_ordering = "none"), file='NUL') coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) Z_SVC_test <- cbind(c(0.1,0.3,0.7),c(0.5,0.2,0.4)) gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all") capture.output( pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, gp_rand_coef_data_pred = Z_SVC_test, cov_pars = c(0.1,1,0.1,0.8,0.15,1.1,0.08), predict_cov_mat = TRUE) , file='NUL') expected_mu <- c(-0.1669209, 1.6166381, 0.2861320) expected_cov <- c(9.643323e-01, 3.536846e-04, -1.783557e-04, 3.536846e-04, 5.155009e-01, 4.554321e-07, -1.783557e-04, 4.554321e-07, 7.701614e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.1,1,0.1,0.8,0.15,1.1,0.08),y=y) expect_lt(abs(nll-149.4422184),1E-5) # Fit model using gradient descent with Nesterov acceleration capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, gp_rand_coef_data = Z_SVC, vecchia_ordering = "none", y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = FALSE, maxit=10)), file='NUL') expected_values <- c(0.3448993, 0.2302483, 0.7981342, 0.4158840, 0.1514441, 0.1074046, 1.1479748, 0.7761315, 0.1032126, 0.1012417, 0.3224399, 0.6417908, 0.1061352, 0.2955655) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-expected_values)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 10) # Prediction capture.output( gp_model <- GPModel(gp_coords = coords, gp_rand_coef_data = Z_SVC, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, vecchia_ordering = "none"), file='NUL') coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) Z_SVC_test <- cbind(c(0.1,0.3,0.7),c(0.5,0.2,0.4)) gp_model$set_prediction_data(vecchia_pred_type = "order_obs_first_cond_all", num_neighbors_pred = 30) pred <- gp_model$predict(y = y, gp_coords_pred = coord_test, gp_rand_coef_data_pred=Z_SVC_test, cov_pars = c(0.1,1,0.1,0.8,0.15,1.1,0.08), predict_cov_mat = TRUE) expected_mu <- c(-0.1688452, 1.6181756, 0.2849745) expected_cov <- c(0.9643376, 0.0000000, 0.0000000, 0.0000000, 0.5155030, 0.0000000, 0.0000000, 0.0000000, 0.7702683) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 30, gp_rand_coef_data = Z_SVC, vecchia_ordering = "none", y = y, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc= FALSE, maxit=5)), file='NUL') expected_values <- c(0.0004631625, 0.2001592837, 1.1329783638, 0.4500150650, 0.1466853248, 0.0745943630, 1.6392349806, 0.7922744312, 0.0565169535, 0.0483411364, 0.4393511638, 0.5909479447, 0.0321612593, 0.1046076251) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-expected_values)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 5) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.1,1,0.1,0.8,0.15,1.1,0.08),y=y) expect_lt(abs(nll-149.4840466),1E-6) }) test_that("Wendland covariance function for Gaussian process model ", { y <- eps + xi capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "wendland", cov_fct_taper_shape = 0, cov_fct_taper_range = 0.1), file='NUL') capture.output( fit(gp_model, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5)) , file='NUL') cov_pars <- c(0.002911765, 0.116338096, 0.993996193, 0.211276385) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 280) # Prediction using given parameters capture.output( gp_model <- GPModel(gp_coords = coords, cov_function = "wendland", cov_fct_taper_shape = 1, cov_fct_taper_range = 2), file='NUL') coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = c(0.02,1.2), predict_cov_mat = TRUE) expected_mu <- c(-0.008405567, 1.493836307, 0.720565199) expected_cov <- c(2.933992e-02, 2.223241e-06, 1.352544e-05, 2.223241e-06, 2.496193e-02, 1.130906e-05, 1.352544e-05, 1.130906e-05, 2.405649e-02) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Prediction of variances only pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = c(0.02,1.2), predict_var = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$var)-expected_cov[c(1,5,9)])),1E-6) # Fisher scoring capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "wendland", cov_fct_taper_shape = 0, cov_fct_taper_range = 0.1, y = y, params = list(optimizer_cov = "fisher_scoring", std_dev = TRUE, use_nesterov_acc = FALSE, delta_rel_conv = 1E-6)), file='NUL') cov_pars <- c(2.946448e-08, 1.599928e-01, 1.391589e+00, 2.933997e-01) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 5) # Evaluate negative log-likelihood nll <- gp_model$neg_log_likelihood(cov_pars=c(0.02,1.2),y=y) expect_lt(abs(nll-136.9508962),1E-6) # Other taper shapes capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "wendland", cov_fct_taper_shape = 1, cov_fct_taper_range = 0.15, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5)), file='NUL') cov_pars <- c(0.0564441, 0.0497191, 0.9921285, 0.1752661) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 19) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = c(0.02,1.2), predict_cov_mat = TRUE) expected_mu <- c(-0.007404038, 1.487424320, 0.200022114) expected_cov <- c(1.113020e+00, -6.424533e-30, -4.186440e-22, -6.424533e-30, 3.522739e-01, 9.018454e-10, -4.186440e-22, 9.018454e-10, 6.092985e-01) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) # Other taper shapes capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "wendland", cov_fct_taper_shape = 2, cov_fct_taper_range = 0.08, y = y, params = list(optimizer_cov = "gradient_descent", std_dev = TRUE, lr_cov = 0.1, use_nesterov_acc = TRUE, acc_rate_cov = 0.5)), file='NUL') cov_pars <- c(0.00327103, 0.06579671, 1.08812978, 0.18151366) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),1E-6) expect_equal(gp_model$get_num_optim_iter(), 187) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, cov_pars = c(0.02,1.2), predict_cov_mat = TRUE) expected_mu <- c(-2.314198e-05, 8.967992e-01, 2.430054e-02) expected_cov <- c(1.2200000, 0.0000000, 0.0000000, 0.0000000, 0.9024792, 0.0000000, 0.0000000, 0.0000000, 1.1887157) expect_lt(sum(abs(pred$mu-expected_mu)),1E-6) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),1E-6) }) test_that("Tapering ", { y <- eps + X%*%beta + xi coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) X_test <- cbind(rep(1,3),c(-0.5,0.2,0.4)) # No tapering capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') cov_pars <- c(0.01621846, 0.07384498, 0.99717680, 0.21704099, 0.09616230, 0.03034715) coef <- c(2.30554610, 0.21565230, 1.89920767, 0.09567547) num_it <- 99 expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), num_it) # Prediction pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE) expected_mu <- c(1.195910242, 4.060125034, 3.15963272) expected_cov <- c(6.304732e-01, 1.313601e-05, 1.008080e-07, 1.313601e-05, 3.524404e-01, 3.699813e-07, 1.008080e-07, 3.699813e-07, 4.277339e-01) expect_lt(sum(abs(pred$mu-expected_mu)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),TOLERANCE_STRICT) # With tapering and very large tapering range capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "tapering", cov_fct_taper_shape = 0, cov_fct_taper_range = 1e6, y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), num_it) # Prediction pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),TOLERANCE_STRICT) # With tapering and smaller tapering range capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "tapering", cov_fct_taper_shape = 0, cov_fct_taper_range = 0.5, y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') cov_pars_tap <- c(0.02593993, 0.07560715, 0.99435221, 0.21816716, 0.17712808, 0.09797175) coef_tap <- c(2.32410488, 0.20610507, 1.89498931, 0.09533541) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_tap)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef_tap)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 75) # Same thing with Matern covariance capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 1.5, y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') cov_pars <- c(0.17369771, 0.07950745, 0.84098718, 0.20889907, 0.08839526, 0.01190858) coef <- c(2.33980860, 0.19481950, 1.88058081, 0.09786326) num_it <- 21 expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), num_it) pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE) expected_mu <- c(1.253044, 4.063322, 3.104536) expected_cov <- c(5.880651e-01, 3.732173e-05, 4.443229e-08, 3.732173e-05, 3.627280e-01, 1.497245e-06, 4.443229e-08, 1.497245e-06, 3.796592e-01) expect_lt(sum(abs(pred$mu-expected_mu)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),TOLERANCE_STRICT) # With tapering and very large tapering range capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 1.5, gp_approx = "tapering", cov_fct_taper_shape = 1, cov_fct_taper_range = 1e6, y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), num_it) pred <- predict(gp_model, gp_coords_pred = coord_test, X_pred = X_test, predict_cov_mat = TRUE) expect_lt(sum(abs(pred$mu-expected_mu)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(pred$cov)-expected_cov)),TOLERANCE_STRICT) # With tapering and smaller tapering range capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "matern", cov_fct_shape = 1.5, gp_approx = "tapering", cov_fct_taper_shape = 1, cov_fct_taper_range = 0.5, y = y, X = X, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') cov_pars_tap <- c(0.21355413, 0.08709305, 0.80448797, 0.20623554, 0.12988850, 0.03404038) coef_tap <- c(2.3533920, 0.1896204, 1.8720682, 0.0988744) expect_lt(sum(abs(as.vector(gp_model$get_cov_pars())-cov_pars_tap)),TOLERANCE_STRICT) expect_lt(sum(abs(as.vector(gp_model$get_coef())-coef_tap)),TOLERANCE_STRICT) expect_equal(gp_model$get_num_optim_iter(), 25) }) test_that("Saving a GPModel and loading from file works ", { y <- eps + xi coord_test <- cbind(c(0.1,0.2,0.7),c(0.9,0.4,0.55)) capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) # Save model to file filename <- tempfile(fileext = ".json") saveGPModel(gp_model, filename = filename) rm(gp_model) # Load from file and make predictions again gp_model_loaded <- loadGPModel(filename = filename) pred_loaded <- predict(gp_model_loaded, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expect_equal(pred$mu, pred_loaded$mu) expect_equal(pred$cov, pred_loaded$cov) # With Vecchia approximation capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 20, vecchia_ordering = "none", y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) # Save model to file filename <- tempfile(fileext = ".json") saveGPModel(gp_model, filename = filename) rm(gp_model) # Load from file and make predictions again capture.output( gp_model_loaded <- loadGPModel(filename = filename), file='NUL') pred_loaded <- predict(gp_model_loaded, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expect_equal(pred$mu, pred_loaded$mu) expect_equal(pred$cov, pred_loaded$cov) # With Vecchia approximation and random ordering capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "vecchia", num_neighbors = 20, vecchia_ordering = "random", y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') gp_model$set_prediction_data(num_neighbors_pred = 50) pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) # Save model to file filename <- tempfile(fileext = ".json") saveGPModel(gp_model, filename = filename) rm(gp_model) # Load from file and make predictions again capture.output( gp_model_loaded <- loadGPModel(filename = filename), file='NUL') gp_model_loaded$set_prediction_data(num_neighbors_pred = 50) pred_loaded <- predict(gp_model_loaded, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expect_equal(pred$mu, pred_loaded$mu) expect_equal(pred$cov, pred_loaded$cov) # With Tapering capture.output( gp_model <- fitGPModel(gp_coords = coords, cov_function = "exponential", gp_approx = "tapering", cov_fct_taper_range = 0.5, cov_fct_taper_shape = 1., y = y, params = DEFAULT_OPTIM_PARAMS_STD), file='NUL') pred <- predict(gp_model, y=y, gp_coords_pred = coord_test, predict_cov_mat = TRUE) # Save model to file filename <- tempfile(fileext = ".json") saveGPModel(gp_model, filename = filename) rm(gp_model) # Load from file and make predictions again capture.output( gp_model_loaded <- loadGPModel(filename = filename), file='NUL') pred_loaded <- predict(gp_model_loaded, gp_coords_pred = coord_test, predict_cov_mat = TRUE) expect_equal(pred$mu, pred_loaded$mu) expect_equal(pred$cov, pred_loaded$cov) }) }