Knapsack +

Download as pdf or txt
Download as pdf or txt
You are on page 1of 49

The Knapsack Problem

15
20 W
20
10

n items with weight wi ∈ N and profit pi ∈ N


Choose a subset x of items
P
Capacity constraint
P i∈x wi ≤ W
wlog assume i wi > W , ∀i : wi < W
P
Maximize profit i∈x pi

– 28. April 2010 1/44


Optimization problem

Set of instances I
Function F that gives for all w ∈ I the set of feasible
solutions F (w )
Goal function g that gives for each s ∈ F (w ) the value g(s)
Optimization goal: Given input w , maximize or minimize the
value g(s) among all s ∈ F (w )
Decision problem: Given w ∈ I and k ∈ N, decide whether
OPT (w ) ≤ k (minimization)
OPT (w ) ≥ k (maximization)
where OPT (w ) is the optimal function value among all
s ∈ F (w ).

– 28. April 2010 2/44


Optimization problem

Set of instances I
Function F that gives for all w ∈ I the set of feasible
solutions F (w )
Goal function g that gives for each s ∈ F (w ) the value g(s)
Optimization goal: Given input w , maximize or minimize the
value g(s) among all s ∈ F (w )
Decision problem: Given w ∈ I and k ∈ N, decide whether
OPT (w ) ≤ k (minimization)
OPT (w ) ≥ k (maximization)
where OPT (w ) is the optimal function value among all
s ∈ F (w ).

– 28. April 2010 2/44


Quality of approximation algorithms

Recall: An approximation algorithm A producing a solution of


value A(w ) on a given input w ∈ I has approximation ratio r iff

A(w )
≤r ∀w ∈ I
OPT (w )

(for maximization problems) or

OPT (w )
≤r ∀w ∈ I
A(w )

(for minimization problems)


How good your approximation algorithm is depends on the
value of r and its running time.

– 28. April 2010 3/44


Negative result

We cannot find a result with bounded difference to the optimal


solution in polynomial time.
Interestingly, the problem remains NP-hard if all items have the
same weight to size ratio!

– 28. April 2010 4/44


Reminder?: Linear Programming

Definition
A linear program with n variables and m constraints is specified
by the following minimization problem
Cost function f (x) = c · x
c is called the cost vector
m constraints of the form ai · x ./i bi where ./i ∈ {≤, ≥, =},
ai ∈ Rn We have

L = x ∈ Rn : ∀1 ≤ i ≤ m : xi ≥ 0 ∧ ai · x ./i bi .

Let aij denote the j-th component of vector ai .

– 28. April 2010 5/44


Reminder?: Linear Programming

Definition
A linear program with n variables and m constraints is specified
by the following minimization problem
Cost function f (x) = c · x
c is called the cost vector
m constraints of the form ai · x ./i bi where ./i ∈ {≤, ≥, =},
ai ∈ Rn We have

L = x ∈ Rn : ∀1 ≤ i ≤ m : xi ≥ 0 ∧ ai · x ./i bi .

Let aij denote the j-th component of vector ai .

– 28. April 2010 6/44


Complexity

Theorem
A linear program can be solved in polynomial time.

Worst case bounds are rather high


The algorithm used in practice (simplex algorithm) might
take exponential worst case time
Reuse is not only possible but almost necessary

– 28. April 2010 7/44


Integer Linear Programming

ILP: Integer Linear Program, A linear program with the


additional constraint that all the xi ∈ Z
Linear Relaxation: Remove the integrality constraints from an
ILP

– 28. April 2010 8/44


Example: The Knapsack Problem

maximize p · x
subject to

w · x ≤ W , xi ∈ {0, 1} for 1 ≤ i ≤ n.

xi = 1 iff item i is put into the knapsack.


0/1 variables are typical for ILPs

– 28. April 2010 9/44


Linear relaxation for the knapsack
problem

maximize p · x
subject to

w · x ≤ W , 0 ≤ xi ≤ 1 for 1 ≤ i ≤ n.

We allow items to be picked “fractionally”


x1 = 1/3 means that 1/3 of item 1 is put into the knapsack
This makes the problem much easier. How would you solve it?

– 28. April 2010 10/44


The Knapsack Problem

15
20 W
20
10

n items with weight wi ∈ N and profit pi ∈ N


Choose a subset x of items
P
Capacity constraint
P i∈x wi ≤ W
wlog assume i wi > W , ∀i : wi < W
P
Maximize profit i∈x pi

– 28. April 2010 11/44


How to Cope with ILPs

− Solving ILPs is NP-hard


+ Powerful modeling language
+ There are generic methods that sometimes work well
+ Many ways to get approximate solutions.
+ The solution of the integer relaxation helps. For example
sometimes we can simply round.

– 28. April 2010 12/44


Linear Time Algorithm for
Linear Relaxation of Knapsack
pi
Classify elements by profit density into B, {k}, S such that
wi
p p pj
∀i ∈ B, j ∈ S : i ≥ k ≥ , and,
wi wk wj
X X
wi ≤ W but wk + wi > W .
i∈B i∈B
10
B 15
 W

 1 P
if i ∈ B
W − i∈B wi k 20
Set xi = w if i = k

 k
0 if i ∈ S
S 20

– 28. April 2010 13/44



1
 P
if i ∈ B
W− wi
xi = wk
i∈B
if i = k


0 if i ∈ S
Lemma
x is the optimal solution of the linear relaxation.

Proof.
Let x∗ denote the optimal solution
w · x∗ = W otherwise increase some xi 10
∀i ∈ B : xi∗ = 1 otherwise B 15
increase xi∗ and decrease some xj∗ for j ∈ {k} ∪ S W
∀j ∈ S : xj∗ = 0 otherwise k 20

decrease xj and increase xk∗
P
W− wi
This only leaves xk = i∈B
wk
S 20

– 28. April 2010 14/44



1
 P
if i ∈ B
W− wi
xi = wk
i∈B
if i = k


0 if i ∈ S
Lemma
x is the optimal solution of the linear relaxation.

Proof.
Let x∗ denote the optimal solution
w · x∗ = W otherwise increase some xi 10
∀i ∈ B : xi∗ = 1 otherwise B 15
increase xi∗ and decrease some xj∗ for j ∈ {k} ∪ S W
∀j ∈ S : xj∗ = 0 otherwise k 20

decrease xj and increase xk∗
P
W− wi
This only leaves xk = i∈B
wk
S 20

– 28. April 2010 14/44



1
 P
if i ∈ B
W− wi
xi = wk
i∈B
if i = k


0 if i ∈ S
Lemma
x is the optimal solution of the linear relaxation.

Proof.
Let x∗ denote the optimal solution
w · x∗ = W otherwise increase some xi 10
∀i ∈ B : xi∗ = 1 otherwise B 15
increase xi∗ and decrease some xj∗ for j ∈ {k} ∪ S W
∀j ∈ S : xj∗ = 0 otherwise k 20

decrease xj and increase xk∗
P
W− wi
This only leaves xk = i∈B
wk
S 20

– 28. April 2010 14/44



1
 P
if i ∈ B
W− wi
xi = wk
i∈B
if i = k


0 if i ∈ S
Lemma
x is the optimal solution of the linear relaxation.

Proof.
Let x∗ denote the optimal solution
w · x∗ = W otherwise increase some xi 10
∀i ∈ B : xi∗ = 1 otherwise B 15
increase xi∗ and decrease some xj∗ for j ∈ {k} ∪ S W
∀j ∈ S : xj∗ = 0 otherwise k 20

decrease xj and increase xk∗
P
W− wi
This only leaves xk = i∈B
wk
S 20

– 28. April 2010 14/44



1
 P
if i ∈ B
W− i∈B wi
xi = wk if i = k


0 if i ∈ S
Lemma
X
opt ≤ xi pi ≤ 2opt
i

Proof.
P
We have i∈B pi ≤ opt. Furthermore,
10
since wk < W , pk ≤ opt. We get B 15 W
X X
opt ≤ xi pi ≤ pi + pk k 20
i i∈B
≤ opt + opt = 2opt
S 20

– 28. April 2010 15/44


Two-approximation of Knapsack


1
 P
if i ∈ B
W− wi
xi = i∈B
wk if i = k 10

 B
0 if i ∈ S 15 W
Exercise: Prove that either B or {k} is a k 20
2-approximation of the (nonrelaxed)
knapsack problem. S 20

– 28. April 2010 16/44


Dynamic Programming
— Building it Piece By Piece

Principle of Optimality
An optimal solution can be viewed as constructed of
optimal solutions for subproblems
Solutions with the same objective values are
interchangeable
Example: Shortest Paths
Any subpath of a shortest path is a shortest path
Shortest subpaths are interchangeable

s u v t

– 28. April 2010 17/44


Dynamic Programming by Capacity
for the Knapsack Problem

Define
P(i, C) = optimal profit from items 1,. . . ,i using capacity ≤ C.

Lemma

∀1 ≤ i ≤ n : P(i, C) = max(P(i − 1, C),


P(i − 1, C − wi ) + pi )

Of course this only holds for C large enough: we must have


C ≥ wi .

– 28. April 2010 18/44


Lemma
∀1 ≤ i ≤ n : P(i, C) = max(P(i − 1, C), P(i − 1, C − wi ) + pi )

Proof.
To prove: P(i, C) ≤ max(P(i − 1, C), P(i − 1, C − wi ) + pi )
Assume the contrary ⇒
∃x that is optimal for the subproblem such that

P(i − 1, C) < p · x ∧ P(i − 1, C − wi ) + pi < p · x

Case xi = 0: x is also feasible for P(i − 1, C). Hence,


P(i − 1, C) ≥ p · x. Contradiction
Case xi = 1: Setting xi = 0 we get a feasible solution x0 for
P(i − 1, C − wi ) with profit p · x0 = p · x − pi . Add
pi . . .

– 28. April 2010 19/44


Computing P(i, C) bottom up:

Procedure knapsack(p, c, n, W )
array P[0 . . . W ] = [0, . . . , 0]
bitarray decision[1 . . . n, 0 . . . W ] = [(0, . . . , 0), . . . , (0, . . . , 0)]
for i := 1 to n do
// invariant: ∀C ∈ {1, . . . , W } : P[C] = P(i − 1, C)
for C := W downto wi do
if P[C − wi ] + pi > P[C] then
P[C] := P[C − wi ] + pi
decision[i, C] := 1

– 28. April 2010 20/44


Recovering a Solution

C := W
array x[1 . . . n]
for i := n downto 1 do
x[i] := decision[i, C]
if x[i] = 1 then C := C − wi
endfor
return x

Analysis:
Time: O(nW ) pseudo-polynomial
Space: W + O(n) words plus Wn bits.

– 28. April 2010 21/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2
3
4

– 28. April 2010 22/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
Entries in table are P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2
3
4
We check each time whether P[C − wi ] + pi > P[C]

– 28. April 2010 23/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
Entries in table are P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2
3
4
We check each time whether P[C − wi ] + pi > P[C]

– 28. April 2010 24/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
Entries in table are P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2 0, (0) 10, (0) 10, (0) 20, (1) 30, (1) 30, (1)
3
4
We check each time whether P[C − wi ] + pi > P[C]

– 28. April 2010 25/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
Entries in table are P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2 0, (0) 10, (0) 10, (0) 20, (1) 30, (1) 30, (1)
3 0, (0) 10, (0) 15, (1) 25, (1) 30, (0) 35, (1)
4
We check each time whether P[C − wi ] + pi > P[C]

– 28. April 2010 26/44


Example: A Knapsack Instance

maximize (10, 20, 15, 20) · x


subject to (1, 3, 2, 4) · x ≤ 5
Entries in table are P(i, C), (decision[i, C])
i \C 0 1 2 3 4 5
0 0 0 0 0 0 0
1 0, (0) 10, (1) 10, (1) 10, (1) 10, (1) 10, (1)
2 0, (0) 10, (0) 10, (0) 20, (1) 30, (1) 30, (1)
3 0, (0) 10, (0) 15, (1) 25, (1) 30, (0) 35, (1)
4 0, (0) 10, (0) 15, (0) 25, (0) 30, (0) 35, (0)
We check each time whether P[C − wi ] + pi > P[C]

– 28. April 2010 27/44


Dynamic Programming by Profit
for the Knapsack Problem

Define
C(i, P) = smallest capacity from items 1,. . . ,i giving profit ≥ P.

Lemma

∀1 ≤ i ≤ n : C(i, P) = min(C(i − 1, P),


C(i − 1, P − pi ) + wi )

– 28. April 2010 28/44


Dynamic Programming by Profit

Let P̂:= bp · x∗ c where x ∗ is the optimal solution of the linear


relaxation.
Thus P̂ is the value (profit) of this solution.
 
Time: O nP̂ pseudo-polynomial
Space: P̂ + O(n) words plus P̂n bits.

– 28. April 2010 29/44


A Faster Algorithm

Dynamic programs are only pseudo-polynomial-time


A polynomial-time solution is not possible (unless P=NP...),
because this problem is NP-hard
However, it would be possible if the numbers in the input were
small (i.e. polynomial in n)
To get a good approximation in polynomial time, we are going
to ignore the least significant bits in the input

– 28. April 2010 30/44


Fully Polynomial Time Approximation
Scheme

Algorithm A is a
(Fully) Polynomial Time Approximation Scheme
minimization
for problem Π if:
maximization
Input: Instance I, error parameter ε
≤ 1+ε
Output Quality: f (x) ( )opt
≥ 1−ε
Time: Polynomial in |I| (and 1/ε)

– 28. April 2010 31/44


Example Bounds

PTAS FPTAS
1
n + 21/ε n2 +
ε
1 1
nlog ε n+ 4
ε
1
nε n/ε
3 ..
n42/ε .
1000/ε ..
n + 22 .
.. ..
. .

– 28. April 2010 32/44


Problem classes

We can classify problems according to the approximation ratios


which they allow.
APX: constant approximation ratio achievable in time
polynomial in n (Metric TSP, Vertex Cover)
PTAS: 1 + ε achievable in time polynomial in n for any
ε > 0 (Euclidean TSP)
FPTAS: 1+ achievable in time polynomial in n and 1/ε for
any > 0 (Knapsack)

– 28. April 2010 33/44


FPTAS → optimal solution

By choosing ε small enough, you can guarantee that the


solution you find is in fact optimal. The running time will depend
on the size of the optimal solution, and will thus again not be
strictly polynomial-time (for all inputs).

– 28. April 2010 34/44


FPTAS for Knapsack

Recall that pi ∈ N for all i!

P:= maxi pi // maximum profit


εP
K := // scaling factor
n 
pi0 := pKi // scale profits
x0 := dynamicProgrammingByProfit(p0 , c, C)
output x0

– 28. April 2010 35/44


FPTAS for Knapsack

16
20 W
21
Recall that pi ∈ N for all i! 11
P:= maxi pi // maximum profit
εP
K := // scaling factor
n 
pi0 := pKi // scale profits
x0 := dynamicProgrammingByProfit(p0 , c, C)
output x0

Example:
ε = 1/3, n = 4, P = 20 → K = 5/3
p = (11, 20, 16, 21) → p 0 = (6, 12, 9, 12) (or p0 = (2, 4, 3, 4))
– 28. April 2010 36/44
Lemma
p · x0 ≥ (1 − ε)opt.

Proof.
Consider the optimal solution x∗ .
X j p k
p · x∗ − K p0 · x∗ = pi − K i
K
i∈x∗
X   pi 
≤ pi − K − 1 = |x ∗ |K ≤ nK ,

K
i∈x

i.e., K p0 · x∗ ≥ p · x∗ − nK . Furthermore,
X j pi k X pi
K p0 · x∗ ≤ K p0 · x0 = K ≤ K = p · x0 .
0
K 0
K
i∈x i∈x

We use that x0 is an optimal solution for the modified problem.

p · x0 ≥K p0 · x∗ ≥ p · x∗ − nK = opt − εP ≥ (1 − ε)opt
– 28. April 2010 37/44
Lemma
p · x0 ≥ (1 − ε)opt.

Proof.
Consider the optimal solution x∗ .
X j p k
p · x∗ − K p0 · x∗ = pi − K i
K
i∈x∗
X   pi 
≤ pi − K − 1 = |x ∗ |K ≤ nK ,

K
i∈x

i.e., K p0 · x∗ ≥ p · x∗ − nK . Furthermore,
X j pi k X pi
K p0 · x∗ ≤ K p0 · x0 = K ≤ K = p · x0 .
0
K 0
K
i∈x i∈x

We use that x0 is an optimal solution for the modified problem.

p · x0 ≥K p0 · x∗ ≥ p · x∗ − nK = opt − εP ≥ (1 − ε)opt
– 28. April 2010 37/44
Lemma
p · x0 ≥ (1 − ε)opt.

Proof.
Consider the optimal solution x∗ .
X p
i

p · x∗ − K p0 · x∗ ≤ pi − K − 1 = |x ∗ |K ≤ nK ,

K
i∈x

i.e., K p0 · x∗ ≥ p · x∗ − nK . Furthermore,
X j pi k X pi
K p0 · x∗ ≤ K p0 · x0 = K ≤ K = p · x0 .
0
K 0
K
i∈x i∈x

Hence,

p · x0 ≥K p0 · x∗ ≥ p · x∗ − nK = opt − εP ≥ (1 − ε)opt

– 28. April 2010 38/44


Lemma

Running time O n3 /ε .

Proof.
The running time ofdynamic
 programming dominates.
Recall that this is O nPb0 where Pb0 = bp0 · x∗ c.
We have
   
0 2 P 2 Pn n3
nP̂ ≤n · (n · max pi ) = n
0 =n ≤ .
K εP ε

– 28. April 2010 39/44


A Faster FPTAS for Knapsack

Simplifying assumptions:
1/ε ∈ N: Otherwise ε:= 1/ d1/εe.
Upper bound P̂ is known: Use linear relaxation to get a quick
2-approximation.
mini pi ≥ εP̂: Treat small profits separately. For these items
greedy works well. (Costs a factor O(log(1/ε))
time.)

– 28. April 2010 40/44


A Faster FPTAS for Knapsack
1
M:= ; K := P̂ε2 = P̂/M
ε2
jp k 
i
pi0 := // pi0 ∈ 1ε , . . . , M
K
value of optimal solution
 was at most P̂, is now M
Define bucketsj Cjk:= i ∈ 1..n : pi0 = j
M
keep only the j lightest (smallest) items from each Cj
do dynamic programming on the remaining items

Lemma
px0 ≥ (1 − ε)opt.

Proof.
Similar as before, note that |x| ≤ 1/ε for any solution.

– 28. April 2010 41/44


Lemma
Running time O(n + Poly(1/ε)).

Proof.

preprocessing time: O(n)


values: M = 1/ε2
XM   XM  
M 1 log(1/ε)
pieces: ≤M ≤ M ln M = O
j j ε2
i=1/ε i=1/ε
 
time dynamic programming: O log(1/ε)
ε4

– 28. April 2010 42/44


The Best Known FPTAS

[Kellerer, Pferschy 04]


( )!
1 log2 1ε
O min n log + ,...
ε ε3

Less buckets Cj (nonuniform)


Sophisticated dynamic programming

– 28. April 2010 43/44


Optimal Algorithm for the Knapsack
Problem

The best work in near linear time for almost all inputs! Both in a
probabilistic and in a practical sense.
[Beier, Vöcking, An Experimental Study of Random Knapsack
Problems, European Symposium on Algorithms, 2004.]
[Kellerer, Pferschy, Pisinger, Knapsack Problems, Springer
2004.]
Main additional tricks:
reduce to core items with good profit density,
Horowitz-Sahni decomposition for dynamic programming

– 28. April 2010 44/44

You might also like