Solving the Knapsack Problem with Groovy
Author: Paul King
Published: 2024-02-09 03:00PM
The knapsack problem is a problem in combinatorial optimization. Given a set of items, each of a given weight and value, determine which items to place into a knapsack (or other container) in order to maximise the value within the knapsack without exceeding a given weight limit.
Case Study
For our case study, we’ll start with placing gems within the knapsack. The gems have the following characteristics:
Index | Weight | Value |
---|---|---|
0 |
1 |
1 |
1 |
2 |
5 |
2 |
3 |
10 |
3 |
5 |
15 |
4 |
6 |
17 |
The gem can either be in the knapsack or not. This is known as the 0/1 knapsack variation of the problem. We’ll look at some other variations later.
Our goal then is to find out which gems we place into the knapsack to maximise the value.
Brute force
The first approach we might take is simply to try all possible combinations, throwing away any which exceed the weight limit, and then finding the maximum value of those that are left.
One way to do this is to generate all possible index combinations. Then, select the ones which satisfy the knapsack weight constraint. Then, find the one which yields the highest value as shown here:
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
var totalValue = { it*.v.sum() }
var withinLimit = { it*.w.sum() <= W }
var asTriple = { [i:it, w:weights[it], v:values[it]] }
var best = weights
.indices
.collect(asTriple)
.subsequences()
.findAll(withinLimit)
.max(totalValue)
println "Total value for capacity $W = ${totalValue(best)}"
println "Gems taken: ${best*.i}"
println "Gem weights: ${best*.w}"
In a bit more detail, we first define three helper closures, totalValue
, withinLimit
,
and asTriple
. Collecting our data into the list of maps (triples) isn’t required
but simplifies subsequent processing. Once we have our data, we find the
subsequences, retain the ones within the weight limit, and then find the maximum value.
Running this script yields the following output:
Total value for capacity 10 = 30 Gems taken: [1, 2, 3] Gem weights: [2, 3, 5]
While this is simple enough, it doesn’t offer many opportunities for optimisation. For a small problem like our case study, optimisation is not important, but for larger problems, the number of combinations may grow exponentially large, and we’ll want to keep that in mind.
Let’s instead, consider a recursive solution which lets us opt out earlier for cases which will exceed the maximum knapsack weight limit:
int solve(int[] w, int[] v, int W) {
knapsack(w, v, v.length, W)
}
int knapsack(int[] w, int[] v, int n, int W) {
if (n <= 0) {
0
} else if (w[n - 1] > W) {
knapsack(w, v, n - 1, W)
} else {
[knapsack(w, v, n - 1, W),
v[n - 1] + knapsack(w, v, n - 1, W - w[n - 1])].max()
}
}
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
[6, 8, 10].each {
println "Total value for capacity $it = ${solve(weights, values, it)}"
}
Here, we consider each gem (for a given stage, gem n
, or index n - 1
).
There are two paths to calculate, one with the gem included,
the other with it excluded. We remember the maximum value returned
from the two paths and continue processing. When a gem is included,
we then solve the smaller problem of fitting the remaining gems into a conceptually
smaller knapsack. If placing a gem into the knapsack causes the weight to exceed
the limit, then we can discard that path from further processing.
It is useful to visualize the above process as a solution tree (shown for capacity 10):
The light red nodes indicate where subsequent processing of the solution tree can be skipped.
We calculate here the result for 3 different knapsack weight limits (6, 8, and 10). The output looks like this:
Total value for capacity 6 = 17 Total value for capacity 8 = 25 Total value for capacity 10 = 30
Instead of having 32 combinations (2^5 for our 5 gems), we’ll only process 11, 16, and 21 combinations where both paths are traversed when the maximum weight limit is 6, 8, and 10 respectively.
More importantly, we’ll have further possibilities for optimisation.
One quick win is to memoize (cache) the results of calling the knapsack
method.
Groovy makes this easy. The only change from above is the addition of the @Memoized
annotation:
@Memoized
int knapsack(int[] w, int[] v, int n, int W) {
if (n <= 0) {
0
} else if (w[n - 1] > W) {
knapsack(w, v, n - 1, W)
} else {
[knapsack(w, v, n - 1, W),
v[n - 1] + knapsack(w, v, n - 1, W - w[n - 1])].max()
}
}
Running this has the same result as before but the number of executions of
the knapsack
method reduces from 44 to 32 (when just calculating for the knapsack of weight limit 10), and from 107 to 49 (when calculating for 6, 8, and 10).
Using Branch and Bound
Another technique often used for optimisation is branch and bound. It’s a special form of the general principle of divide and conquer; solving a big problem by turning it into smaller problems.
Divide and conquer is similar to what we did for the recursive solution above, but with branch and bound, we perform smarter elimination. Before processing the children of a branch, the branch is checked against upper and lower estimated bounds of some optimal solution. Processing for a given path is terminated if we can determine that heading down that path can’t possibly be better than heading done some alternative path. For the knapsack problem, we can work out those bounds by finding the optimal "greedy" solution if we were allowed to use fractional quantities of a given knapsack item. We’ll look at fractional quantities as the last example in this blog. It turns out we can calculate them very efficiently.
First, we’ll create an Item
record for holding our weights and values.
record Item(int weight, int value) {}
Next, we’ll create a Node
to hold information about the current status
of a candidate solution at a particular point in our solution tree:
@Canonical
class Node {
int level, value, weight
public int bound
static Node next(Node parent) {
new Node(parent.level + 1, parent.value, parent.weight)
}
}
Next, the bound
method calculates the weight using the
optimal (greedy) algorithm. This would require us to allow fractional
parts of gems to be accurate, but in our case, we are just using it
as a bound. Think of estimating with best and worst case kept in mind.
int bound(Node u, int n, int W, List<Item> items) {
if (u.weight >= W)
return 0
int valueBound = u.value
int j = u.level + 1
int totalWeight = u.weight
while (j < n && totalWeight + items[j].weight <= W) {
totalWeight += items[j].weight
valueBound += items[j].value
j++
}
if (j < n)
valueBound += (int) ((W - totalWeight) * items[j].value / items[j].weight)
valueBound
}
Now, our knapsack method will sort the gems according to most value per weight and then process through them keeping track of the bound at each step.
int knapsack(int W, List<Item> items, int n) {
items.sort { it.value / it.weight }
var q = new PriorityQueue<>((a, b) -> b.bound <=> a.bound)
Node u, v
q.offer(new Node(-1, 0, 0))
int bestValue = 0
while (q) {
u = q.poll()
if (u.level == n - 1)
continue
else
v = Node.next(u)
v.weight += items[v.level].weight
v.value += items[v.level].value
if (v.weight <= W && v.value > bestValue)
bestValue = v.value
v.bound = bound(v, n, W, items)
if (v.bound > bestValue)
q.offer(v)
v = Node.next(u)
v.bound = bound(v, n, W, items)
if (v.bound > bestValue)
q.offer(v)
}
bestValue
}
int W = 10
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
var items = values.indices.collect {
new Item(weights[it], values[it])
}
println "Total value for capacity $W = ${knapsack(W, items, values.length)}"
Which has this visualization (for capacity 10):
We should note that as well as discarding the infeasible paths which exceed the weight limit (light red), we now also discard unfruitful paths (light purple) which our bound value tells us can never exceed some alternative best path we already know about.
It has the following output:
Total value for capacity 10 = 30
Using Dynamic Programming
You can think of
dynamic programming
as a special case of the
branch and bound technique. It can also be thought of as similar
to the memoization we looked at earlier. In this case, rather
than Groovy providing us with the cache, we track it ourselves
in the dp
array:
int solve(int[] v, int[] w, int W) {
knapsack(new Integer[v.length][W+1], v, w, W, 0)
}
int knapsack(Integer[][] dp, int[] v, int[] w, int W, int n) {
if (W <= 0 || n >= v.length) {
return 0
}
if (dp[n][W]) {
return dp[n][W]
}
int tryN = w[n] > W ? 0 : v[n] + knapsack(dp, v, w, W - w[n], n + 1)
int skipN = knapsack(dp, v, w, W, n + 1)
dp[n][W] = max(tryN, skipN)
}
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
[6, 8, 10].each {
println "Total value for capacity $it = ${solve(values, weights, it)}"
}
To solve the knapsack problem, we break it into smaller pieces. If we have already cached the solution to the smaller piece, we use the cached value. Because of the way we have structured our problem, we are actually sharing the cached results for the different knapsack sizes.
When we run this script, it produces the following output:
Total value for capacity 6 = 17 Total value for capacity 8 = 25 Total value for capacity 10 = 30
Like with most things, we have options when using dynamic programming. Here is an alternative variant which keeps a second array tracking the gems taken:
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
int N = values.length
int[][] dp = new int[N + 1][W + 1]
boolean[][] sol = new boolean[N + 1][W + 1]
for (int n = 1; n <= N; n++) {
for (int w = 1; w <= W; w++) {
int skipN = dp[n - 1][w]
int tryN = weights[n - 1] > w ? 0 : values[n - 1] + dp[n - 1][w - weights[n - 1]]
dp[n][w] = max(skipN, tryN)
sol[n][w] = tryN > skipN
}
}
println "Total value for capacity $W = ${dp[N][W]}"
def taken = []
for (int i = N, j = W; j > 0; i--) {
if (sol[i][j]) {
taken << i - 1
j -= weights[i - 1]
}
}
println "Gems taken: $taken"
If just the final value is what we want to work out, our earlier variant will be very slightly faster and use less memory. If keeping track of the gems taken is important, then this variant would be one way to go.
It produces the following output:
Total value for capacity 10 = 30 Gems taken: [3, 2, 1]
Using BitSets
When using dynamic programming, our 2D dp
array could
use significant memory for large problems. In such cases,
we could use a bitset instead of the array like this:
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
var N = weights.size()
var nums = 0L..<(1L << N)
var totalValue = nums
.collect{ BitSet.valueOf(it) }
.findAll{ mask -> mask.stream().map(idx -> weights[idx]).reduce(0, Integer::sum) <= W }
.collect{ mask -> nums.indices.sum{ idx -> mask[idx] ? values[idx] : 0 } }
.max()
println "Total value for capacity $W = $totalValue"
Here we are using the bitset to track the gem combinations in a candidate solution. This might seem a little unusual but does the job in our case.
It has the following output:
Total value for capacity 10 = 30
Most often, bitsets are used not just to save memory but because for certain problems we can perform operations on entire bitsets. You can think of this as bulk operations with free parallelism when compared to performing such operations on say arrays of booleans.
We can show a simple example of this kind of operation
by adding a preliminary optimisation step. We’ll use a simple
trick which can find all possible sums of a set of numbers
using bit shifting. If the maximum weight (10 in this example)
isn’t in the list of possible sums - we find that out using the
maskW
constant, then we discard the combination.
For simplicity, we’ve kept the example simple here, but realise that this
crude optimisation has the possibility of ruling out valid candidates.
The maximum value could, in theory, be for weight 8 or 9 for instance.
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
var N = weights.size()
var maskW = 1L << W
var nums = 0L..<(1L << N)
var totalValue = nums
.collect{ BitSet.valueOf(it) }
.findAll{ mask -> mask.stream().reduce(1) {a, b -> a | a << weights[b] } & maskW }
.findAll{ mask -> mask.stream().map(idx -> weights[idx]).reduce(0, Integer::sum) <= W }
.collect{ mask -> nums.indices.sum{ idx -> mask[idx] ? values[idx] : 0 } }
.max()
println "Total value for capacity $W = $totalValue"
Which has the same output for this case study, so luckily, our crude optimisation step didn’t reject the best solution.
In ths example, we just shifted an integer. Depending on the
particular problem, we might wan tto instead shift a long or bitset.
Groovy 5 adds support for <<
(left shift),>>
(right shift),
and >>>
(right shift unsigned) operators for bitsets.
This kind of functionality will make working on such
problems even nicer with Groovy.
Using Constraint Programming
Another technique we could consider is constraint programming. Here we define some constraints and let a solver find solutions for us. Here we have used the Choco solver.
We thought we would also spice up the example and illustrate what is known as the bounded knapsack problem. Instead of either taking the gem or leaving it out, we now consider gems to be readily available commodities and the weight and value would apply to all gems of a particular type.
In general, we can take as many gems of a particular type as we want (this would be unbounded), or as we’ll do here take as many gems of a particular type up to some bound.
In our example, we’ll set the bound for our solver’s
domain variables to be between 0
and W
.
We could easily set the upper bound to be 1
and
we’d be back to the 0/1 knapsack problem.
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
int unbounded = 100000
var counts = new IntVar[values.length]
var found = false
new Model('KnapsackProblem').with {
counts.indices.each(i -> counts[i] = intVar("count$i", 0, W))
scalar(counts, weights, '<=', W).post()
var total = intVar("Total value for capacity $W (unbounded)", 0, unbounded)
scalar(counts, values, '=', total).post()
setObjective(MAXIMIZE, total)
while (solver.solve()) {
found = true
println "$total, $counts"
}
}
if (!found) println 'No solution'
We keep an array counts
of solver variables,
apply numerous constraints on the variables,
tell the solver to maximise the total
variable.
When we run this script, it produces the following output:
Total value for capacity 10 (unbounded) = 25, [count0 = 0, count1 = 5, count2 = 0, count3 = 0, count4 = 0] Total value for capacity 10 (unbounded) = 30, [count0 = 0, count1 = 2, count2 = 2, count3 = 0, count4 = 0] Total value for capacity 10 (unbounded) = 31, [count0 = 1, count1 = 0, count2 = 3, count3 = 0, count4 = 0]
Here, the solver is finding solutions which satisfy the problem as we have specified it, and then tries to find better solutions. We should note that because we are allowing more than one of each gem, it isn’t surprising that our answer (31) is higher than our previous best answer (30).
If we didn’t want to receive all solutions, we can simply ask for the best solution, or provide the solver with better search hints for the problem, to arrive at the best answer earlier.
As it turns out, the set of constraints we set in place here to
solve the knapsack problem, are so common, that Choco has a
built-in knapsack
method which sets multiple constraints
for us. We could use it as follows:
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
int W = 10
int unbounded = 100000
var counts = new IntVar[values.length]
var found = false
new Model('KnapsackProblem').with {
counts.indices.each(i -> counts[i] = intVar("count$i", 0, W))
var totalValue = intVar("Total value for capacity $W (unbounded)", 0, unbounded)
var totalWeight = intVar("Total weight taken", 0, W)
knapsack(counts, totalWeight, totalValue, weights, values).post()
setObjective(MAXIMIZE, totalValue)
while (solver.solve()) {
found = true
println "$totalValue, $totalWeight, $counts"
}
}
if (!found) println 'No solution'
Which when run has this output:
Total value for capacity 10 (unbounded) = 30, Total weight taken = 10, [count0 = 0, count1 = 2, count2 = 2, count3 = 0, count4 = 0] Total value for capacity 10 (unbounded) = 31, Total weight taken = 10, [count0 = 1, count1 = 0, count2 = 3, count3 = 0, count4 = 0]
Using OrTools
Other libraries also have built-in solvers for knapsack. Here is another implementation using the OrTools library:
Loader.loadNativeLibraries()
var solver = new KnapsackSolver(KNAPSACK_MULTIDIMENSION_BRANCH_AND_BOUND_SOLVER, "knapsack")
long[] values = [1, 5, 10, 15, 17]
long[][] weights = [[1, 2, 3, 5, 6]]
long[] capacities = [10]
solver.init(values, weights, capacities)
long computedValue = solver.solve()
println "Total value for capacity ${capacities[0]} = " + computedValue
var packedItems = []
var packedWeights = []
int totalWeight = 0
values.indices.each { i ->
if (solver.bestSolutionContains(i)) {
packedItems << i
packedWeights << weights[0][i]
totalWeight += weights[0][i]
}
}
println "Actual weight: $totalWeight"
println "Gems taken: $packedItems"
println "Gem weights: $packedWeights"
Which when run has this output:
Total value for capacity 10 = 30 Actual weight: 10 Gems taken: [1, 2, 3] Gem weights: [2, 3, 5]
Using Jenetics
We can also use Genetic Algorithms to find a solution. When creating solutions based on genetic algorithms, a series of steps take place which mimic evolution in nature. We typically start with some random guesses, which we regard as the first generation of children. We have functions to test the fitness of individuals, processes to select a new generation of children based on the current generation, steps for mutation and so forth.
We’ll start by creating a record to track our weights and values:
record Item(int weight, int value) implements Serializable {}
We’ll create a Knapsack
class to keep track of state
and provide our fitness function:
class Knapsack implements Problem<ISeq<Item>, BitGene, Integer> {
private final Codec<ISeq<Item>, BitGene> codec
private final double knapsackSize
Knapsack(ISeq<Item> items, int knapsackSize) {
codec = Codecs.ofSubSet(items)
this.knapsackSize = knapsackSize
}
@Override
Function<ISeq<Item>, Integer> fitness() {
(items) -> {
var sum = items.inject(new Item(0, 0)) { acc, next ->
new Item(acc.weight + next.weight, acc.value + next.value)
}
sum.weight <= knapsackSize ? sum.value : 0
}
}
@Override
Codec<ISeq<Item>, BitGene> codec() { codec }
}
We create our genetic algorithm engine and configure it work out how the next generation will be selected, and what mutations, if any, might be used to provide random alterations.
We’ll also create a log
function to output some
information as our various generations are being produced.
Here is the script:
int W = 10
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
var items = [weights, values].transpose().collect { w, v -> new Item(w, v) }
var iSeq = items.stream().collect(ISeq.toISeq())
var knapsack = new Knapsack(iSeq, W)
var engine = Engine.builder(knapsack)
.populationSize(8)
.survivorsSelector(new TournamentSelector<>(3))
.offspringSelector(new RouletteWheelSelector<>())
.alterers(
new Mutator<>(0.1),
new SinglePointCrossover<>(0.2),
new MultiPointCrossover<>(0.1))
.build()
var log = { EvolutionResult er ->
var avg = er.population().average{ it.fitness() }
var best = er.bestFitness()
printf "avg = %5.2f, best = %d%n", avg, best
}
var best = engine.stream()
.limit(bySteadyFitness(8))
.limit(30)
.peek(log)
.collect(toBestPhenotype())
println best
Which when run produces this output:
avg = 18.88, best = 23 avg = 21.00, best = 25 avg = 22.00, best = 25 avg = 22.63, best = 25 avg = 25.63, best = 30 avg = 27.50, best = 30 avg = 27.63, best = 30 avg = 24.38, best = 30 avg = 22.50, best = 30 avg = 26.25, best = 30 avg = 30.00, best = 30 avg = 30.00, best = 30 [00001110] -> 30
Since it is random, subsequent runs may produce different results:
avg = 16.75, best = 27 avg = 17.13, best = 23 avg = 18.00, best = 23 avg = 21.38, best = 27 avg = 24.00, best = 27 avg = 24.88, best = 27 avg = 22.50, best = 27 avg = 26.88, best = 27 [00010100] -> 27
As we can see here, we aren’t guaranteed to get the optimal solution with all genetic algorithm problems.
What we should notice is that while the process involves various random aspects, and we might sometimes actually kill off best individuals, if we have configured our algorithm correctly, we should see that the average and best fitness values go up over time.
Using Greedy selection
The final example we will look at is the "optimal" or "greedy" solution. Here we take the items in order of best value/weight. If we allowed a fractional value greater than 1, we’d just use the first item in this sorted list, but here we’ll have a maximum of 1 for any item.
For this problem, instead of gems, which might indeed be hard to split (or at least split and not devalue significantly), you might instead think of exotic spices, or some other valuable which can be readily divided.
Here is the code:
int[] values = [1, 5, 10, 15, 17] // Gem values
int[] weights = [1, 2, 3, 5, 6] // Weights
var ratios = values.indices.collect { values[it] / weights[it] }.withIndex().sort { -it[0] }
int W = 10
Map<Integer, BigDecimal> taken = [:]
var remaining = W
while (remaining && ratios) {
var next = ratios.head()
var index = next[1]
if (remaining >= weights[index]) {
taken[index] = 1
remaining -= weights[index]
} else {
taken[index] = remaining / weights[index]
break
}
ratios = ratios.tail()
}
var total = taken.collect{ index, qty -> values[index] * qty }.sum()
println taken
printf 'Total value for capacity %d (with fractions) = %6.3f%n', W, total
Which when run has this output:
[2:1, 3:1, 4:0.3333333333] Total value for capacity 10 (with fractions) = 30.667
It is not unexpected that we can obtain a value greater than 30 when allowing fractional amounts of the valuables.
Further information
-
Choco Solver: home page, a different knapsack example
-
OR-Tools: home page, a different knapsack example
Conclusion
We have seen how to solve the knapsack problem in Groovy using several approaches. In the references, there are even more exotic algorithms which can be used to solve the knapsack problem. If you have a great way to solve the knapsack problem using Groovy, let us know and we can add it to this blog!