Table of content

    Adjustable robust optimization with discrete uncertainty > DisruptionFLP

    gap = function(LB, UB) {
      return(100 * abs(LB - UB) / (1e-10 + abs(UB)))
    }
    key = c("instance", "gamma")
    time_limit = 3600
    raw_benders = read.csv("./results_benders_DisruptionFLP.csv", header = FALSE)
    colnames(raw_benders) <- c("instance", "gamma", "status", "reason", "objective", "time", "nodes", "rel_gap", "abs_gap", "n_generated_cuts")
    paged_table(raw_benders)

    Unifying format

    To make our study easier, we start by unifying the format of each dataset. To do so, we transform our data to obtain it in the following format:

    • instance: the instance filename ;
    • n_facilities: the number of facilities ;
    • n_customers: the number of customers ;
    • gamma: the value of \(\Gamma\) ;
    • objective: the best objective value found (feasible) ;
    • time: the computation time spent solving the instance.

    We thus introduce two functions read_csv_benders and read_csv_ccg which first reads an input file for the corresponding approach and returns the associated formatted table.

    Before that, we first introduce a helper function parse_instance_properties which takes a list of instances as input and returns a table containing, for each instance, the number of knapsacks, the number of items and the value for alpha extracted from the instance file name.

    library(stringr)
    
    parse_instance_properties = function (instances) {
      
      parsed = t(apply(as.matrix(instances), 1, function(str)  str_extract_all(str, regex("([0-9]+)"))[[1]]))
      
      result = data.frame(instances, as.double(parsed[,1]), as.double(parsed[,2]), as.double(parsed[,3]))
      colnames(result) = c("instance", "n_facilities", "n_customers", "ratio")
      
      return (result)
      
    }

    read_csv_benders

    The `read_csv_benders`` function is given as follows.

    read_csv_benders = function(filename) {
      
      # Read raw results
      raw_results = read.csv(filename, header = FALSE)
      colnames(raw_results) <- c("instance", "gamma", "status", "reason", "objective", "time", "nodes", "rel_gap", "abs_gap", "n_generated_cuts")
      
      # Fix unsolved instances to TIME_LIMIT
      if (sum(raw_results$time >= 3600) > 0) {
        raw_results[raw_results$time >= 3600,]$time = 3600
      }
      
      # Extract properties from instance file names
      properties = parse_instance_properties(raw_results$instance)
      
      # Build result data frame
      result = data.frame(
          properties$instance, 
          properties$n_facilities,
          properties$n_customers,
          properties$ratio,
          raw_results$gamma,
          raw_results$objective,
          raw_results$time,
          raw_results$nodes,
          raw_results$rel_gap * 100,
          "B&C"
        )
      colnames(result) = c("instance", "n_facilities", "n_customers", "ratio", "gamma", "objective", "time", "nodes", "gap", "solver")
      
      result = result[result$ratio != 150,]
      
      return (result)
      
    }

    We can then read CSV files coming from the benders approach.

    benders = read_csv_benders("./results_benders_DisruptionFLP.csv")

    read_csv_ccg

    The `read_csv_ccg`` function is given as follows.

    read_csv_ccg = function (filename) {
      
      # Read raw results
      raw_results = read.csv(filename, header = FALSE)
      colnames(raw_results) <- c("instance", "gamma", "iter", "LB", "UB", "time", "inner_iter_1", "inner_iter_2")
      
      # Fix unsolved instances to TIME_LIMIT
      if (sum(raw_results$time >= 3600) > 0) {
        raw_results[raw_results$time >= 3600,]$time <- 3600
      }
      
      # Extract properties from instance file names
      properties = parse_instance_properties(raw_results$instance)
      
      # Build result data frame
      result = data.frame(
          properties$instance, 
          properties$n_facilities,
          properties$n_customers,
          properties$ratio,
          raw_results$gamma,
          raw_results$UB,
          raw_results$time,
          NA,
          gap(raw_results$LB, raw_results$UB),
          "CCG"
        )
      colnames(result) = c("instance", "n_facilities", "n_customers", "ratio", "gamma", "objective", "time", "nodes", "gap", "solver")
      
      raw_results[abs(raw_results$UB) < 1e-6 & abs(raw_results$LB) < 1e-6,]$gap = 0
      
      result = result[result$ratio != 150,]
      
      return (result)
      
    }

    Then, we read the results obtained by the CCG approach as follows.

    ccg = read_csv_ccg("./results_ccg_DisruptionFLP.csv")
    ## Warning in `[<-.data.frame`(`*tmp*`, abs(raw_results$UB) < 0.000001 &
    ## abs(raw_results$LB) < : provided 11 variables to replace 10 variables

    Checking

    compare = function(a, b) {
      A = a[a$time < 3600,c("instance", "gamma", "objective", "time")]
      B = b[b$time < 3600,c("instance", "gamma", "objective", "time")]
      
      merged = merge(A, B, by = c("instance", "gamma"))
      
      filter = gap(merged$objective.x, merged$objective.y) > 1e-3
      if (sum( filter ) > 0) {
        print("For some instances, the two methods do not agree.")
        paged_table( merged[filter,] )
      }
    }
    
    compare(benders, ccg)

    Performance profile

    all_results = rbind(benders, ccg)
    
    solvers = unique(all_results$solver)
    
    colors = cbind(as.data.frame(solvers), rainbow(length(solvers)))
    colnames(colors) = c("solver", "color")
    performance_profile = function (dataset, xlim = NULL, main = "Performance profile") {
    
      times = spread(dataset[,c("instance", "gamma", "solver", "time")], key = solver, value = time)
      times$time.best = apply(times[,-c(1,2)], 1, FUN = min)
      
      #times = na.omit(times)
      #print("WARNING omitting NA")
      
      ratios = times[,-ncol(times)][,-c(1,2)] / times$time.best
      colnames(ratios) = paste0(colnames(ratios), ".ratio")
      
      worst_ratio = max(ratios)
      
      times = cbind(times, ratios)
      
      for (solver in solvers) {
        time_limit_filter = times[,solver] >= time_limit
        if ( sum(time_limit_filter) > 0 ) {
          times[time_limit_filter, paste0(solver, ".ratio")] = worst_ratio
        }
      }
      
      if (is.null(xlim)) {
        xlim = c(1, worst_ratio)
      }
      
      #par(mar = c(5,4,4,8))
      
      using_colors = NULL
      using_types = NULL
      
      last_ecdf = NULL
      
      index = 1
      for (solver in solvers) {
        
        plot_function = if (index == 1) plot else lines
        
        profile = ecdf(times[,paste0(solver, ".ratio")])
        
        using_color =  colors[colors$solver == solver,2]
        using_colors = rbind(using_colors, using_color)
        using_type = "solid"
        if (using_color == "#00FFFFFF") {
          using_type = "dashed"
        }
        using_types = rbind(using_types, using_type)
        
        plot_function(profile, xlim = xlim, ylim = c(0,1), lty = using_type, cex = 0, col = using_color, main = "", xlab = "", ylab = "")
        
        if (is.null(last_ecdf)) {
          last_ecdf = profile
        } else {
          p = seq(0, 1, length.out = 50000)
          df = data.frame(
            quantile(last_ecdf, probs = p),
            quantile(profile, probs = p)
          )
          colnames(df) = solvers
          print( head( df[df["CCG"] > df["B&C"],], 1 ))
        }
        
        index = index + 1
      }
      
      # Set the plot title
      title(main = main,
            xlab = "Performance ratio",
            ylab = "ECDF")
      
      # Set the plot legend
      legend(
        "bottomright",
        #inset=c(-.35, 0),
        legend = solvers,
        lty = using_types,
        col = using_colors,
        #cex = .5,
        #xpd = TRUE,
        bty = "n"
      )
    }
    {
      performance_profile(all_results)
      performance_profile(all_results, xlim = c(1, 100), main = "Performance profile")
    }

    ##                   B&C     CCG
    ## 72.99145983% 6.533554 6.53648

    ##                   B&C     CCG
    ## 72.99145983% 6.533554 6.53648
    #for (gamma in unique(all_results$gamma)) {
    #  performance_profile(all_results[all_results$gamma == gamma,], main = paste("With gamma =", gamma))
    #}
    
    for (ratio in unique(all_results$ratio)) {
      performance_profile(all_results[all_results$ratio == ratio,], xlim = c(1, 70), main = "")
    }

    ##                   B&C      CCG
    ## 74.57349147% 22.18092 22.24441

    ##                   B&C      CCG
    ## 65.64931299% 2.468363 2.469409

    Summary tables

    group_by = c("ratio", "gamma", "n_facilities", "n_customers")
    str_group_by = c("$\\mu$", "$\\Gamma$", "$|V_1|$", "$|V_2|$")

    Computing summary times by group

    We start by calling summary on each group. This will compute, for each group, the minimum, 1st quantile, median, mean, 3rd quantile and maximum execution time. This is done in the following function which takes as parameter the formatted table with all results.

    compute_summary_by_group = function(data) {
      
      # We summarize only the solved instances
      data = data[data$time < 3600,]
      
      # We aggregate by `group_by` using `summary`
      result = aggregate(data$time, by = data[,group_by], summary)
      
      # Here, we call data.frame recursively to flatten `result` (summary does create a nested structure)
      result = do.call(data.frame, result)
      
      # Then, we set column names
      colnames(result) = c(group_by, "min", "1st_quantile", "median", "mean", "3rd_quantile", "max")
      
      return (result)
    }

    We then call this function on each approach.

    summary_ccg = compute_summary_by_group(ccg)
    summary_benders = compute_summary_by_group(benders)

    Computing number of solved instances by group

    We then count the number of instances which could not be solved to optimality within each group. This is done in the following function.

    compute_solved_by_group = function(data) {
    
      # Aggregate groups using `sum` over the filter returning 1 iff the time limit is reached
      result = aggregate(data$time < 3600, by = data[,group_by], sum)
      
      # Set column and row names
      colnames(result) = c(group_by, "solved")
      rownames(result) = NULL
      
      return (result)
      
    }

    Again, we call this function on each approach.

    solved_ccg = compute_solved_by_group(ccg)
    solved_benders = compute_solved_by_group(benders)

    Computing average gap of unsolved instances by group

    We then compute the average final gap over those instances which could not be solved to optimality within each group. This is done in the following function.

    compute_gap_by_group = function(data) {
      
      # Aggregate groups using `sum` over the filter returning 1 iff the time limit is reached
      result = aggregate(data$gap, by = data[,group_by], mean)
      
      # Set column and row names
      colnames(result) = c(group_by, "gap")
      rownames(result) = NULL
      
      return (result)
      
    }

    Again, we call this function on each approach.

    gap_ccg = compute_gap_by_group(ccg)
    gap_benders = compute_gap_by_group(benders)

    Computing average number of nodes instances by group

    We then compute the average final gap over those instances which could not be solved to optimality within each group. This is done in the following function.

    compute_nodes_by_group = function(data) {
    
      only_solved = data[data$time < 3600,]
      
      # Aggregate groups using `sum` over the filter returning 1 iff the time limit is reached
      result = aggregate(only_solved$node, by = only_solved[,group_by], mean)
      
      # Set column and row names
      colnames(result) = c(group_by, "node")
      rownames(result) = NULL
      
      return (result)
      
    }

    Again, we call this function on each approach.

    nodes_ccg = compute_nodes_by_group(ccg)
    nodes_benders = compute_nodes_by_group(benders)

    Computing number of instances by group

    Finally, we also count the number of instances which was tried by each approach. (Note that when experiments are done, they should all be equal).

    compute_instances_by_group = function(data) {
      
      # Call length on each group to count the number of instances
      result = aggregate(data$instance, by = data[,group_by], length)
      
      # Set row and column names
      colnames(result) = c(group_by, "total")
      rownames(result) = NULL
      
      return (result)
      
    }

    Let’s call it on each approach.

    total_ccg = compute_instances_by_group(ccg)
    total_benders = compute_instances_by_group(benders)

    Summary table

    Final result

    We are now ready to build our summary table. In what follows, we introduce a helper function to add the useful columns of a by-group result (i.e., total_<APPROACH>, unsolved_<APPROACH> or summary_<APPROACH>) to the main table, called Table.

    # Helper function
    add_columns = function(table, data, suffix) {
      result = merge(table, data, by = group_by, all = TRUE, suffixes = c("", paste(".", suffix, sep = "")))
      return (result)
    }
    
    # We list the useful columns of the by-group results
    column_names = c("total", "solved", "min", "1st_quantile", "median", "mean", "3rd_quantile", "max", "gap", "node")
    
    # We create an empty data frame with column names. This is used to
    # (1) create the placeholder for the group identifiers
    # (2) force R to rename each added column by appending a suffix to it
    Table = data.frame(matrix(ncol = length(group_by) + length(column_names), nrow = 0))
    colnames(Table) = c(group_by, column_names)
    
    # We add by-group totals
    Table = add_columns(Table, total_benders, "B&C")
    Table = add_columns(Table, total_ccg, "CCG")
    
    # We add by-group solved
    Table = add_columns(Table, solved_benders, "B&C")
    Table = add_columns(Table, solved_ccg, "CCG")
    
    # We add by-group summaries
    Table = add_columns(Table, summary_benders, "B&C")
    Table = add_columns(Table, summary_ccg, "CCG")
    
    # We add by-group gap
    Table = add_columns(Table, gap_benders, "B&C")
    Table = add_columns(Table, gap_ccg, "CCG")
    
    # We add by-group gap
    Table = add_columns(Table, nodes_benders, "B&C")
    
    # Finally, we can remove the "fake" columns we introduced to foce R renaming new columns
    Table = Table[,!names(Table) %in% column_names]

    The resulting table therefore contains a merge of all by-group data and is drawn hereafter.

    To better summarize our data, we make a new table out of this table by selecting interesting columns. This is done as follows.

    Table1 = Table[,c(
      group_by,
      "solved.B&C",
      "solved.CCG",
      "mean.B&C",
      "mean.CCG",
      "gap.B&C",
      "gap.CCG",
      "node.B&C"
      )]
    Solved
    Time
    Gap
    Node
    \(\mu\) \(\Gamma\) \(&amp;#124;V_1&amp;#124;\) \(&amp;#124;V_2&amp;#124;\) B&C CCG B&C CCG B&C CCG B&C
    200 2 10 20 10 10 13.61 4.98 0.00 0.00 19
    200 2 10 30 10 10 64.50 106.81 0.00 10.00 36
    200 2 10 40 10 8 254.90 40.25 0.00 7.95 27
    200 2 10 50 10 10 1058.59 106.21 0.00 0.00 30
    200 2 15 20 10 9 122.09 134.87 0.00 12.46 63
    200 2 15 30 10 6 742.82 553.81 0.00 5.19 60
    200 2 15 40 6 7 1509.53 92.54 1.93 4.96 82
    200 2 15 50 5 6 1882.22 560.80 2.70 5.93 50
    200 3 10 20 10 10 22.74 8.46 0.00 30.91 31
    200 3 10 30 10 10 75.81 37.17 0.00 0.00 38
    200 3 10 40 10 10 213.18 95.80 0.00 20.00 36
    200 3 10 50 8 8 658.77 451.47 27.12 2.04 34
    200 3 15 20 10 7 218.62 69.42 0.00 32.40 88
    200 3 15 30 10 4 1320.35 177.90 0.00 19.18 55
    200 3 15 40 4 7 1353.41 468.41 5.55 2.57 58
    200 3 15 50 1 3 3572.38 856.32 8.89 39.34 51
    200 4 10 20 10 10 34.25 3.38 0.00 30.01 84
    200 4 10 30 10 10 121.93 37.88 0.00 25.94 162
    200 4 10 40 10 10 360.44 198.41 0.00 72.07 165
    200 4 10 50 8 9 1000.25 306.56 7.72 14.66 107
    200 4 15 20 10 6 452.02 246.01 0.00 9985652759414.99 496
    200 4 15 30 7 3 806.49 147.91 13.49 13452069676498.95 72
    200 4 15 40 3 4 1210.74 1280.90 7.00 13.95 78
    200 4 15 50 0 2 NA 1241.74 39.03 108.12 NA
    300 2 10 20 10 10 13.42 12.83 0.00 0.00 47
    300 2 10 30 10 9 18.50 251.17 0.00 1.98 32
    300 2 10 40 10 10 15.92 70.54 0.00 0.00 30
    300 2 10 50 10 7 158.93 94.25 0.00 19.77 47
    300 2 15 20 10 9 69.06 110.83 0.00 2.42 72
    300 2 15 30 10 7 102.49 406.04 0.00 9.19 33
    300 2 15 40 9 5 324.90 142.36 0.07 13.47 87
    300 2 15 50 10 9 250.11 445.83 0.00 4.72 45
    300 3 10 20 10 9 29.79 37.91 0.00 14295287681237.00 99
    300 3 10 30 10 10 104.54 49.70 0.00 0.00 80
    300 3 10 40 10 10 45.13 36.49 0.00 10.00 65
    300 3 10 50 10 7 230.14 254.23 0.00 62.61 66
    300 3 15 20 10 8 126.99 109.51 0.00 5.53 89
    300 3 15 30 10 6 359.67 264.51 0.00 16.85 60
    300 3 15 40 9 2 1093.62 190.89 0.85 23.41 100
    300 3 15 50 10 6 908.34 768.34 0.00 13.62 65
    300 4 10 20 10 10 41.74 7.63 0.00 7.10 227
    300 4 10 30 10 10 185.98 31.53 0.00 40.08 380
    300 4 10 40 10 10 56.94 18.36 0.00 10.00 110
    300 4 10 50 10 8 607.93 610.69 0.00 2.16 117
    300 4 15 20 10 9 376.20 576.29 0.00 2.47 675
    300 4 15 30 10 5 1195.95 182.24 0.00 63.28 207
    300 4 15 40 5 2 1875.86 562.71 11.97 30.50 113
    300 4 15 50 4 2 1190.49 704.55 21.23 41.09 77

    Solved instances

    plot_solved = function (dataset, main = "Solved instances") {
      data = dataset[, c("solved.B&C", "solved.CCG")]
      rownames(data) = paste0("(", dataset$n_facilities, ",", dataset$n_customers, ")")
      data = t(as.matrix(data))
      barplot( data , beside = TRUE, col = colors$color, main = main, ylim = c(0, 10), cex.names=.8)
      legend("bottomleft",  legend = c("B&C", "CCG"), fill = colors$color)
    }
    for (ratio in unique(Table$ratio)) {
      for (gamma in unique(Table$gamma)) {
        plot_solved(Table[Table$gamma == gamma & Table$ratio == ratio,], main = paste("Solved instances for Gamma =", gamma))
      }
    }

    Partial disruptions (\(K > 1\))

    In this section, we study how B&C performs when increasing the number of unknown coefficients. This leads to the partial disruption FLP application (Note that \(K\) has been replaced by \(R\) in the paper).

    We start by parsing the results stored in the “./resutls_PartialDisruptionFLP.csv” file.

    read_csv_benders_partial = function(filename) {
      
      # Read raw results
      raw_results = read.csv(filename, header = FALSE)
      colnames(raw_results) <- c("instance", "gamma", "status", "reason", "objective", "time", "nodes", "LB", "UB", "n_generated_cuts", "k")
      
      # Fix unsolved instances to TIME_LIMIT
      if (sum(raw_results$time >= 3600) > 0) {
        raw_results[raw_results$time >= 3600,]$time = 3600
      }
      
      # Extract properties from instance file names
      properties = parse_instance_properties(raw_results$instance)
      
      # Build result data frame
      result = data.frame(
          properties$instance, 
          properties$n_facilities,
          properties$n_customers,
          properties$ratio,
          raw_results$gamma,
          raw_results$objective,
          raw_results$time,
          raw_results$nodes,
          "B&C",
          raw_results$k
        )
      colnames(result) = c("instance", "n_facilities", "n_customers", "ratio", "gamma", "objective", "time", "nodes", "solver", "k")
      
      return (result)
      
    }
    
    raw_benders_partial = read_csv_benders_partial("./results_PartialDisruptionFLP.csv")

    Then, we transform these raw data to obtain, for each instance, the computational time required to solve the instance for different values of \(K\).

    by_k = spread(raw_benders_partial[, c("instance", "n_facilities", "n_customers", "ratio", "gamma", "time", "k")], key = k, value = time)
    rownames(by_k) = NULL

    Then, we first make sure that each instance have been solved for \(K = 1, 2, 3\) and \(4\) by printing rows where NA appear.

    paged_table(by_k[!complete.cases(by_k),])

    Median computation times

    by_k = na.omit(by_k)
    partial_mean = aggregate(by_k[,c("1", "2", "3", "4")], by = by_k[,c("n_facilities", "n_customers", "gamma")], median)
    Median time
    \(&amp;#124;V_1&amp;#124;\) \(&amp;#124;V_2&amp;#124;\) \(\Gamma\) 1 2 3 4
    10 20 2 13.39 45.74 96.05 164.51
    15 20 2 65.77 212.10 573.99 1404.95
    10 30 2 14.91 64.47 139.69 283.84
    15 30 2 123.60 620.90 1886.20 3600.00
    10 40 2 17.57 66.29 220.63 506.31
    10 50 2 41.43 500.92 1529.26 3233.98

    Graphical representation

    plot_evolution = function(dataset, main = "Evolution of computational times depending on K") {
      
      #dataset$type = paste0("(", dataset$n_facilities, ",", dataset$n_customers, "), Gamma = ", dataset$gamma)
      dataset$type = paste0("(", dataset$n_facilities, ",", dataset$n_customers, ")")
      
      #longest = max(dataset[ ,"4"])
      
      longest = 3600
      
      colors = rainbow(length(dataset$type))
      
      #par(mar = c(5,4,4,8))
      
      index = 1
      for (type in dataset$type) {
        
        plot_function = if (index == 1) plot else lines
        x = c(1,2,3,4)
        y = as.vector( t(dataset[dataset$type == type, c("1", "2", "3", "4")]) )
        
        if (index == 1) {
          plot_function(x, y, xlim = c(1, 4), ylim = c(0,longest), lty = index, type = "l", cex = 2, col = "black", main = "", xlab = "", ylab = "", axes = FALSE)
        } else {
          plot_function(x, y, xlim = c(1, 4), ylim = c(0,longest), lty = index, type = "l", cex = 2, col = "black", main = "", xlab = "", ylab = "")
        }
        
        index = index + 1
      }
      
      box()
      axis(side = 1, at = seq(from = 1, to = 4, by = 1))
      axis(side = 2, at = seq(from = 0, to = longest, by = 500))
      
      # Set the plot title
      title(main = main,
            ylab = "Median time (s)",
            xlab = "R")
      
      # Set the plot legend
      legend(
        "topleft",
        #inset=c(-.35, 0),
        legend = dataset$type,
        lty = c(1:index),
        col = "black",
        cex = 1,
        xpd = TRUE,
        bty = "n"
      )
    
    }
    plot_evolution(partial_mean[partial_mean$gamma == 2 & partial_mean$n_facilities == 10,], main = "")

    Solved instances

    Here, we count the number of solved instances for each value of \(K\).

    data_increasing_R = raw_benders_partial[raw_benders_partial$gamma == 2 & raw_benders_partial$n_facilities == 10,]
    solved = aggregate(data_increasing_R$time < 3600, by = list(data_increasing_R$k, data_increasing_R$n_customers), sum)
    colnames(solved) = c("K", "n.customers", "n.solved")
    K n.customers n.solved
    1 20 10
    2 20 10
    3 20 10
    4 20 10
    1 30 10
    2 30 10
    3 30 10
    4 30 9
    1 40 10
    2 40 9
    3 40 10
    4 40 9
    1 50 10
    2 50 10
    3 50 7
    4 50 5
    K n.solved
    1 40
    2 39
    3 37
    4 33

    This document is automatically generated after every git push action on the public repository hlefebvr/hlefebvr.github.io using rmarkdown and Github Actions. This ensures the reproducibility of our data manipulation. The last compilation was performed on the 20/11/24 21:13:03.