I am not sure this is possible with reasonable effort, because the tuner does not know what parameters are being set by it. Maybe using something like what I propose in mlr-org/paradox#225 would make this possible, because it would be a $trafo() that knows what the parameters are that are being set. Alternatively the user could somehow inform the AutoTuner what parameters are set by the tuner and should not be set by the user.

Example of a trafo that sets the maxdepth or the cp parameter, but that does not make the fact that it refers to maxdepth or cp obvious to the outside:

> ps = ParamSet$new(list(ParamLgl$new("x")))
> ps$trafo = function(x, param_set)
+   if (!"x" %in% names(x)) x else  # circumvent bug in EvalPerf
+   if (x$x) list(cp = 0.5) else
+   list(maxdepth = 1)

(Tuning happens the following way:)

> tune = TunerRandomSearch$new(
+   pe = PerfEval$new(task = "iris", learner = "classif.rpart",
+     resampling = "cv", measure = "classif.ce", param_set = ps),
+   terminator = TerminatorEvaluations$new(3))$tune()