RLCS: small code rewrites

Small updates of the code

As announced, I’m doing some code rewriting, just because…

Function Factory

So I had 2 or three of these:

inc_match_count <- function(M_pop) { ## All versions
  lapply(M_pop, \(x) {
    x$match_count <- x$match_count + 1
    x
  })
}

inc_correct_count <- function(C_pop) { ## SL Specific
  lapply(C_pop, \(x) {
    x$correct_count <- x$correct_count + 1
    x
  })
}

inc_action_count <- function(A_pop) { ## RL Specific
  lapply(A_pop, \(x) {
    x$action_count <- x$action_count + 1
    x
  })
}

Now it looks like so:

## Function factory to increase parameter counts
inc_param_count <- function(param) {
  param <- as.name(param)
  function(pop) {
    lapply(pop, \(x) {
      x[[param]] <- x[[param]] + 1
      x
    })
  }
}

inc_match_count <- inc_param_count("match_count")
inc_correct_count <- inc_param_count("correct_count")
inc_action_count <- inc_param_count("action_count")

As I do not intend to expose these functions in the future “packaged” version, well, this is pretty safe.

Object for Hyperparameters

And as also announced, I cleaned up a bit the code to “call training”. Now using an object for the hyperparameters of the algorithm, which has its own defaults, makes things a bit cleaner.

source("run_params/datamining_examples_recommended_hyperparameters_v001.R")

basic_hyperparameters <- RLCS_hyperparameters(
  wildcard_prob = wildcard_prob,
  rd_trigger = rd_trigger,
  mutation_probability = mutation_probability,
  parents_selection_mode = parents_selection_mode,
  tournament_pressure = tournament_pressure,
  n_epochs = n_epochs,
  deletion_trigger = deletion_trigger,
  deletion_threshold = deletion_threshold
)

## It makes it more readable here:
example_lcs <- rlcs_train(train_environment, basic_hyperparameters)

And given the number of parameters, it’s even more readable of course if you take the defaults:

default_lcs_hyperparameters <- RLCS_hyperparameters()
example_lcs <- rlcs_train(train_environment, default_lcs_hyperparameters)

Using an object might also help with adding error controls… Probably, that’s not implemented yet. Also, it’s a valid approach to clean code, as these variables are all part of the same concept of hyper-parameters for the algorithm, meaning they do belong together.

And as a note here, “defaults” are seldom all good for an LCS, so you should probably overwrite some of the parameters depending on each problem you look at…

For comparison, I saved a version of an old call, and it’s just that it received too many parameters:

lcs_res <- rlcs_meta_train(train_environment,
                           1, ## Warmup with just one epoch
                           wildcard_prob,
                           rd_trigger,
                           parents_selection_mode,
                           mutation_probability,
                           tournament_pressure,
                           deletion_trigger) ## Deletion won't be triggered

Resources

Clean Code, by R. C. Martin (“Uncle Bob”)

Advanced R, by H. Wickham