Random Seed and Reproducibility - Tutorial (RLT)

Ruoqing Zhu

Last Updated: May 18, 2026

Overview

This page explains how to control randomness in your analyses with base R’s set.seed().

Where to set the seed

At the very beginning of your analysis (recommended). This covers synthetic data generation and internal resampling randomness in RLT(). Or immediately before model fitting if your data are fixed and only modeling randomness matters.

Prerequisites — See Get Started.

Demonstration — same seed, same results (regression)

We run the same pipeline twice with the same seed and compare outputs.

# ---------- Run A (seed = 1) ----------
set.seed(1)

# Small dataset (~100 obs)
trainn <- 80; testn <- 20; n <- trainn + testn; p <- 10
X1 <- matrix(rnorm(n * (p/2)), n, p/2)
X2 <- matrix(as.integer(runif(n * (p/2)) * 3), n, p/2)  # integers 0,1,2
X_numeric <- data.frame(X1, X2)

y <- 1 + rowSums(X_numeric[, 2:6]) +
  2 * (X_numeric[, p/2 + 1] %in% c(1, 2)) + rnorm(n)

X <- X_numeric
X[, (p/2 + 1):p] <- lapply(X[, (p/2 + 1):p], as.factor)

trainX <- X[1:trainn, ]; trainY <- y[1:trainn]
testX  <- X[(trainn + 1):(trainn + testn), ]; testY <- y[(trainn + 1):(trainn + testn)]

# Fit
library(RLT)
ntrees <- 200; ncores <- 1
nmin <- 5; mtry <- p/2; samplereplace <- TRUE; sampleprob <- 0.80
rule <- "best"; nsplit <- ifelse(rule == "best", 0, 3); importance <- TRUE

fit_A <- RLT(
  trainX, trainY, model = "regression",
  ntrees = ntrees, mtry = mtry, nmin = nmin,
  resample.prob = sampleprob, split.gen = rule,
  resample.replace = samplereplace,
  nsplit = nsplit, importance = importance,
  param.control = list(alpha = 0),
  ncores = ncores, verbose = FALSE
)
pred_A <- predict(fit_A, testX, ncores = ncores)

mse_train_A <- mean((fit_A$Prediction - trainY)^2)
mse_test_A  <- mean((pred_A$Prediction - testY)^2)

# ---------- Run B (same seed = 1) ----------
set.seed(1)

# Recreate the same data and pipeline
trainn <- 80; testn <- 20; n <- trainn + testn; p <- 10
X1 <- matrix(rnorm(n * (p/2)), n, p/2)
X2 <- matrix(as.integer(runif(n * (p/2)) * 3), n, p/2)
X_numeric <- data.frame(X1, X2)
y <- 1 + rowSums(X_numeric[, 2:6]) +
  2 * (X_numeric[, p/2 + 1] %in% c(1, 2)) + rnorm(n)

X <- X_numeric
X[, (p/2 + 1):p] <- lapply(X[, (p/2 + 1):p], as.factor)

trainX <- X[1:trainn, ]; trainY <- y[1:trainn]
testX  <- X[(trainn + 1):(trainn + testn), ]; testY <- y[(trainn + 1):(trainn + testn)]

fit_B <- RLT(
  trainX, trainY, model = "regression",
  ntrees = ntrees, mtry = mtry, nmin = nmin,
  resample.prob = sampleprob, split.gen = rule,
  resample.replace = samplereplace,
  nsplit = nsplit, importance = importance,
  param.control = list(alpha = 0),
  ncores = ncores, verbose = FALSE
)
pred_B <- predict(fit_B, testX, ncores = ncores)

mse_train_B <- mean((fit_B$Prediction - trainY)^2)
mse_test_B  <- mean((pred_B$Prediction - testY)^2)

# ---------- Summary for same-seed runs ----------
list(
  A_Train_MSE = round(mse_train_A, 6),
  A_Test_MSE  = round(mse_test_A, 6),
  B_Train_MSE = round(mse_train_B, 6),
  B_Test_MSE  = round(mse_test_B, 6),
  SameSeed_Predictions_Identical = isTRUE(all.equal(pred_A$Prediction, pred_B$Prediction))
)
## $A_Train_MSE
## [1] 3.456843
## 
## $A_Test_MSE
## [1] 4.086383
## 
## $B_Train_MSE
## [1] 3.456843
## 
## $B_Test_MSE
## [1] 4.086383
## 
## $SameSeed_Predictions_Identical
## [1] TRUE

Demonstration — different seed, potentially different results

Now we change the seed and rerun the same pipeline once.

# ---------- Run C (seed = 2) ----------
set.seed(2)

trainn <- 80; testn <- 20; n <- trainn + testn; p <- 10
X1 <- matrix(rnorm(n * (p/2)), n, p/2)
X2 <- matrix(as.integer(runif(n * (p/2)) * 3), n, p/2)
X_numeric <- data.frame(X1, X2)
y <- 1 + rowSums(X_numeric[, 2:6]) +
  2 * (X_numeric[, p/2 + 1] %in% c(1, 2)) + rnorm(n)

X <- X_numeric
X[, (p/2 + 1):p] <- lapply(X[, (p/2 + 1):p], as.factor)

trainX <- X[1:trainn, ]; trainY <- y[1:trainn]
testX  <- X[(trainn + 1):(trainn + testn), ]; testY <- y[(trainn + 1):(trainn + testn)]

fit_C <- RLT(
  trainX, trainY, model = "regression",
  ntrees = ntrees, mtry = mtry, nmin = nmin,
  resample.prob = sampleprob, split.gen = rule,
  resample.replace = samplereplace,
  nsplit = nsplit, importance = importance,
  param.control = list(alpha = 0),
  ncores = ncores, verbose = FALSE
)
pred_C <- predict(fit_C, testX, ncores = ncores)

mse_train_C <- mean((fit_C$Prediction - trainY)^2)
mse_test_C  <- mean((pred_C$Prediction - testY)^2)

list(
  C_Train_MSE = round(mse_train_C, 6),
  C_Test_MSE  = round(mse_test_C, 6),
  DiffSeed_Predictions_EqualTo_RunA = isTRUE(all.equal(pred_C$Prediction, pred_A$Prediction))
)
## $C_Train_MSE
## [1] 3.323691
## 
## $C_Test_MSE
## [1] 2.306646
## 
## $DiffSeed_Predictions_EqualTo_RunA
## [1] FALSE

Tips

• Choose any integer you like for the seed; the specific value doesn’t matter—consistency does.
• Keep one set.seed() near the top of your script to make the whole workflow reproducible.
• The same pattern works for classification and survival: place set.seed() before data simulation (if any) and before RLT().