The issue explained

You define a GraphLearner that inside somewhere has a learner. When you define the Autotuner you provide the search_space of the learner not of the learner inside the larger GraphLearner.

The difference is that for the learner, the parameters that need tuning are defined as alpha and lamdba. Inside the GraphLearner they are defined as surv.glmnet.alpha and surv.glmnet.lambda. This triggers warnings as many lambdas are actually fitted (pretty much the search_space is not used at all in your case I think). You can see that if in your Autotuner you just used the learner, then things would work normally.

This is more general: the GraphLearner constructs <pipeop_id>.<arg_name> to be able to differentiate between parameters of the different pipeops.

Solution(s)

1. Suggested: Define the search_space with the learner (and when the GraphLearner gets constructed, the prefix of the parameters is automatically added)

learner = lrn("surv.glmnet")
learner$param_set$set_values(.values = list(
  alpha = to_tune(0, 1),
  lambda = to_tune(p_dbl(0.001, 0.1, logscale = TRUE))
))

Note that in this case you DO NOT need to use the search_space argument in AutoTuner.

2. Manually define the search space with the suffixes directly given that you don’t change the id = surv.glmnet of the learner, ie:

search_space = ps(
  surv.glmnet.alpha  = p_dbl(lower = 0, upper = 1),   
  surv.glmnet.lambda = p_dbl(lower = 0.0001, upper = 0.1, logscale = TRUE)
)

Suggestions

• You can simplify the pipeline with the colapply as the same operation is applied to all columns, see examples
• Whenever you want just a simple train/test split, do use:

# simple train/test split
part = partition(task)
at$train(task, row_ids = part$train)

• Use the sugar function to construct the autotuner:

at = auto_tuner(
  learner = pipeline, # better name => grlrn, it has the `learner` inside with "solution No 1" above
  resampling = resampling,
  measure = measure,
  tuner = tuner,
  term_evals = 20
)

The revised code included below is working fine. Your suggestions were very helpful. Note that I corrected learner statement. Now it reads learner = lrn("surv.glmnet")

    # Required libraries
    library(mlr3)
    library(mlr3proba)
    library(mlr3learners)
    library(mlr3extralearners)  # added
    library(mlr3tuning)
    library(mlr3pipelines)
    library(survival)
    library(splines)
    library(data.table)
    library(glmnet)

Simulate data

    set.seed(123)
    n = 100
    p = 3
    X = matrix(rnorm(n * p), nrow = n, ncol = p)
    time = rexp(n, rate = 1)
    status = sample(0:1, n, replace = TRUE)
    df = as.data.frame(X)
    df$time <- time
    df$status <- status

Define a survival task

    task <- TaskSurv$new("survival_task", backend = df, time = 
      "time", event = "status")

Load the learner

    learner <- lrn("surv.glmnet")
    learner$param_set$set_values(.values = list(
      alpha  = to_tune(0, 1),
      lambda = to_tune(p_dbl(0.00001, 1, logscale = TRUE))
))

Splines

    apply_splines <- function(x) {
     as.data.table(splines::ns(x, df = 3))
    }

    # Define  grlrn for applying splines transformation
    grlrn0 <- 
     po("colapply", id = "spline_all", 
        applicator = apply_splines, 
        affect_columns = selector_type("numeric")) %>>%
     po("learner", learner)
     grlrn <- GraphLearner$new(grlrn0)

Tuning

    # Resampling strategy for tuning
    resampling <- rsmp("cv", folds = 5)

    # Performance measure for survival analysis
    measure <- msr("surv.cindex")
    
    # Create the tuner
    tuner <- tnr("grid_search", resolution = 10)
    
    # Define the AutoTuner
    at <- auto_tuner(
      learner = grlrn, 
      resampling = resampling,
      measure = measure, 
      ### search_space = search_space, # omitted
      terminator = trm("evals", n_evals = 50),
      tuner = tuner
     )
     
     # Simple train/test split
     part = partition(task)
     at$train(task, row_ids = part$train)
     at$model
     at$tuning_result