Skip to main content

Scala Version of Approximation Algorithm for Knapsack Problem for Apache Spark

This is the Scala version of the approximation algorithm for the knapsack problem using Apache Spark.

I ran this on a local setup, so it may require modification if you are using something like a Databricks environment. Also you will likely need to setup your Scala environment.

All the code for this is at GitHub

First, let's import all the libraries we need.

import org.apache.spark._
import org.apache.spark.rdd.RDD
import org.apache.spark.SparkConf
import org.apache.spark.SparkContext._
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.functions.sum

We'll define this object knapsack, although it could be more specific for what this is doing, it's good enough for this simple test.

object knapsack {

Again, we'll define the knapsack approximation algorithm, expecting a dataframe with the profits and weights, as well as W, a total weight.

  def knapsackApprox(knapsackDF: DataFrame, W: Double): DataFrame = {

Calculate the ratios of profit over weight, and sort them high to low ratio. Discard any weights that are already larger than the max knapsack size, W.

    val ratioDF = knapsackDF.withColumn("ratio", knapsackDF("values") / knapsackDF("weights"))
    val newRatioDF = (ratioDF
      .filter(ratioDF("weights") <= W)

Now we'll use SQL to add up all the partial sums of weights. A window function is another way this could work with SQL. This will tell us what can fit in the knapsack, and remember these are sorted by profit to weight ratio, high to low.

    val partialSumWeightsDF = spark.sql("SELECT item, weights, values, ratio, sum(weights) OVER (ORDER BY ratio desc) as partSumWeights FROM tempTable")
    val partialSumWeightsFilteredDF = (
        .filter(partialSumWeightsDF("partSumWeights") <= W)

And now return this new Dataframe, which will have only the objects that fit.


So this will return the greedy solution, which is fast and easy use parallelism, but is not optimal. Parallel solutions to optimal knapsack algorithms, are often not as simple, but this was a good way to test out Spark using Scala.

And here is the test code, which is pretty self explanatory, the Github is some work in progress and I've some clean up to do.

import org.apache.spark.mllib.random.RandomRDDs._
import scala.collection.mutable.ListBuffer<- ------------------------------------------="" -="" 0.3="" 0.6="" 10.0="" 1="" 5.="" a="" alue="" and="" approximate="" approximation="" call="" countresult="" create="" data:="" data="" dataframe.="" dataframe="" display="" eights="" elected="" elements:="" elements="" end="" find="" for="" function="" greedy="" item="" item_="" k.tostring="""" knapresults="knapsack.knapsackApprox(knapsackData," knapsack.="" knapsack="""" knapsackdata="sc.parallelize(knapsackDataList).toDF(" knapsackdatalist="knapsackDataListBuffer.toList" knapsackdatalistbuffer="" make="" maximum="" n="" of="" original="" ount:="" pre="" println="" r.nextdouble="" r="" random="" results="" riginal="" s="" selected="" show="" size="" start="" test="" the="" to="" total:="" totals.="" totals="" val="" value="" values="""" valuesresult="knapResults.agg(sum(" w="" weight.="" weight="" weights="""" weightsresult="knapResults.agg(sum(" with="">
import org.apache.spark.mllib.random.RandomRDDs._
import scala.collection.mutable.ListBuffer

// Knapsack problem size.
val N = 10

// Random
val r = scala.util.Random

// Setup sample data for knapsack.
val knapsackDataListBuffer = ListBuffer[(String, Double, Double)]()
for (k <- 1 to N) {
  knapsackDataListBuffer += (("item_" + k.toString, r.nextDouble() * 10.0, r.nextDouble() * 10.0))
val knapsackDataList = knapsackDataListBuffer.toList

// Make a Dataframe with item(s), weight(s), and value(s) for the knapsack.
val knapsackData = sc.parallelize(knapsackDataList).toDF("item", "weights", "values")

// Display the original data
println("Original Data:")

// Create a random maximum weight
val start = N * 0.3
val end = N * 0.6
val W = (math.random * (end - start) + start)

// Show the weight.
println("W: ")

// Call the knapsack greedy approximation function, with data and size 5.
val knapResults = knapsack.knapsackApprox(knapsackData, W)

// Show the results Dataframe.
println("Selected Elements:")

// Find the totals.
val valuesResult = knapResults.agg(sum("values"))
val weightsResult = knapResults.agg(sum("weights"))
val countResult = knapResults.count()

// Show totals for selected elements of knapsack.
println("Value Total:")
println("Weights Total:")

And that is it, just create some random items, call the knapsackApprox(knapsackData, W) function, and print out the results. Note, I summed it outside of the main knapsack routine, which just finds the objects that satisfy the problem. Next tasks are: clean up the code for Scala, convert to window function, and complete the Java version.

Popular posts from this blog

Apache Spark Knapsack Approximation Algorithm in Python

The code shown below computes an approximation algorithm, greedy heuristic, for the 0-1 knapsack problem in Apache Spark. Having worked with parallel dynamic programming algorithms a good amount, wanted to see what this would look like in Spark.

The Github code repo. for the Knapsack approximation algorithms is here, and it includes a Scala solution. The work on a Java version is in progress at time of this writing.

Below we have the code that computes the solution that fits within the knapsack W for a set of items each with it's own weight and profit value. We look to maximize the final sum of selected items profits while not exceeding the total possible weight, W.

First we import some spark libraries into Python.

# Knapsack 0-1 function weights, values and size-capacity. from pyspark.sql import SparkSession from pyspark.sql.functions import lit from pyspark.sql.functions import col from pyspark.sql.functions import sum
Now define the function, which will take a Spark Dataframe w…