R/alg_genetic.R

In comeetie/greed: Clustering and Model Selection with the Integrated Classification Likelihood

#' @useDynLib greed
#' @importFrom Rcpp sourceCpp
#' @importFrom future %<-%
#' @name %<-%
NULL

#' @importFrom future %seed%
#' @name %seed%
NULL

#' @include models_classes.R fit_classes.R tools_cleanpath.R
#' @import Matrix
NULL

genetic <- function(model, alg, data, K, verbose = FALSE) {
  init_f <- function(C) {
    fit_greed(model, data, C, "none")
  }
  train.hist <- data.frame(generation = c(), icl = c(), K = c())


  solutions <- listenv::listenv()
  # first generation of solutions
  pop_size <- alg@pop_size
  for (i in 1:pop_size) {
    cli <- sample_cl(model, data, K)
    cli <- as.numeric(factor(cli))
    solutions[[i]] %<-% fit_greed(model, data, cli, type = "none") %seed% TRUE
  }
  solutions <- as.list(solutions)
  icls <- sapply(solutions, function(s) {
    s@icl
  })
  # check for errors
  solutions <- solutions[!is.nan(icls)]
  icls <- icls[!is.nan(icls)]

  nbgen <- 1


  while (max(icls) - min(icls) > 0 && nbgen < alg@nb_max_gen) {
    train.hist <- rbind(train.hist, data.frame(generation = nbgen, icl = icls, K = sapply(solutions, function(s) {
      max(s@cl)
    })))

    ii <- order(icls)
    Nsel <- round(alg@pop_size * alg@sel_frac)
    ii <- ii[(alg@pop_size - Nsel):alg@pop_size]
    icls <- icls[ii]
    solutions <- solutions[ii]
    bres <- solutions[[order(icls, decreasing = TRUE)[1]]]
    new_solutions <- listenv::listenv()
    children <- listenv::listenv()

    for (i in 1:(alg@pop_size - 1)) {
      ip <- 1:Nsel
      i1 <- sample(ip, 1, prob = ip)
      i2 <- sample(ip[-i1], 1, prob = ip[-i1])
      children[[i]] <- cross_over(solutions[[i1]], solutions[[i2]], init_f)
    }
    children <- as.list(children)
    for (i in 1:(alg@pop_size - 1)) {
      if (stats::runif(1) < alg@prob_mutation) {
        new_solutions[[i]] %<-% fit_greed(model, data, children[[i]]@cl, "swap", 1) %seed% TRUE
      } else {
        new_solutions[[i]] <- children[[i]]
      }
    }
    solutions <- c(bres, as.list(new_solutions))

    icls <- sapply(solutions, function(s) {
      s@icl
    })
    nbgen <- nbgen + 1
  }

  train.hist <- rbind(train.hist, data.frame(generation = nbgen, icl = icls, K = sapply(solutions, function(s) {
    max(s@cl)
  })))

  # best solution
  res <- solutions[[order(icls, decreasing = TRUE)[1]]]
  # compute merge path

  res <- fit_greed(model, data, res@cl, "both")
  path <- fit_greed_path(data, res)
  # clean the resuts (compute, merge tree,...)

  path <- cleanpath(path)
  # store train history
  path@train_hist <- train.hist
  path
}


cross_over <- function(sol1, sol2, init_f) {
  cl1 <- sol1@cl
  K1 <- sol1@K
  lk1 <- 1:K1
  cl2 <- sol2@cl
  K2 <- sol2@K
  lk2 <- 1:K2
  C <- rep(0, length(cl1))
  K <- 0
  while (sum(C == 0) > 0) {
    if (K1 > 0 && stats::runif(1) > 0.5) {
      kt <- sample(K1, 1, 1)
      k <- lk1[kt]
      K1 <- K1 - 1
      lk1 <- lk1[-kt]
      ink <- (cl1 == k & C == 0)
      if (sum(ink) > 0) {
        K <- K + 1
        C[ink] <- K
      }
    } else if (K2 > 0) {
      kt <- sample(K2, 1, 1)
      k <- lk2[kt]
      K2 <- K2 - 1
      lk2 <- lk2[-kt]
      ink <- (cl2 == k & C == 0)
      if (sum(ink) > 0) {
        K <- K + 1
        C[ink] <- K
      }
    }
  }
  init_f(C)
}