Dynamic Programming

Data Structures and Algorithms – COMS21103

Dynamic Programming

Largest Empty Square and Weighted Interval Scheduling

Benjamin Sach

The name

Dynamic Programming is an approach to algorithm design. . .

why does it sound like an alternative to Agile Software Development?

The name



- Richard Bellman invented Dynamic programming around 1950

a ‘program’ referred to finding an optimal schedule or programme of activities

Serious answer:

The name



- Richard Bellman invented Dynamic programming around 1950

“The 1950s were not good years for mathematical research. We had a very interesting

gentleman in Washington named Wilson. He was Secretary of Defense, and he actually had a

pathological fear and hatred of the word, research... His face would suffuse, he would turn red,

and he would get violent if people used the term, research, in his presence. You can imagine

how he felt, then, about the term, mathematical... I thought dynamic programming was a good

name. It was something not even a Congressman could object to.”

a ‘program’ referred to finding an optimal schedule or programme of activities

Serious answer:

Real answer:

- Richard Bellman

What problems can Dynamic Programming solve?

• Longest Common Subsequence

• Edit Distance

• Text justification

• Seam Carving

• Solving the Towers of Hanoi

• Predicting cricket scores

• Assembly Line Scheduling

• Matrix Chain Multiplication

• Playing Tetris perfectly

• Dynamic Time Warping

• Finding optimal Binary Search Trees

• The Travelling Salesman Problem

• Knapsack

(used heavily in Bioinformatics for DNA similarity)

(Google this later, it’s really awesome)

(though still slowly)

(when you know the likely frequencies of searches)

(though still slowly)

(used heavily in Bioinformatics for sequence alignment)

(used extensively in computer vision)

and loads of other problems

Introduction

Dynamic programming is recursion without repetition

Dynamic programming is a technique for finding efficient algorithms for problems which

can be broken down into simpler, overlapping subproblems.

in other words. . .

The basic idea:

1. Find a recursive formula for the problem

(typically this is the hard bit)

3. Speed it up by storing the solutions to subproblems(to avoid recomputing the same thing over and over)

4. Derive an iterative algorithm by solving the subproblems in a good order

(iterative algorithms are often better in practice, easier to analyse and prettier)

- in terms of answers to subproblems.

2. Write down a naive recursive algorithm(typically this algorithm will take exponential time)

Part one

Largest Empty Square


Problem Given an n× n monochrome image, find the largest empty square.i.e. without any black pixels

n

n



n

n

an empty square



n

n



n

n

largest empty square

1. Find a recursive formula

To find a recursive formulation of this problem, consider the following fact:

Any m×m square of pixels, S is empty

The bottom right pixel of S is empty and

The three (m− 1)× (m− 1) squares in the

top left, top right and bottom left of S are empty

if and only if







Proof: (by picture)

if and only if







Proof: (by picture)

If S is empty then all four are empty

if and only if







Proof: (by picture)

if and only if







Proof: (by picture)

If all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?


if and only if







Proof: (by picture)

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

empty If all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

empty If all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?


if and only if







Proof: (by picture)

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?

empty


if and only if







Proof: (by picture) empty

????

? ? ? ?

????

? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

????

? ? ? ? ?


if and only if







Proof: (by picture)

????

? ? ? ?

????

? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

????

? ? ? ? ?


if and only if







Proof: (by picture)

????

? ? ? ?

????

? ? ? ?

? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ? ?? ? ? ?

????

? ? ? ? ?

empty


if and only if







Proof: (by picture) empty

? ? ? ?? ? ? ?? ? ? ?? ? ? ? ?


if and only if







Proof: (by picture)

? ? ? ?? ? ? ?? ? ? ?? ? ? ? ?


if and only if







Proof: (by picture)

? ? ? ?? ? ? ?? ? ? ?? ? ? ? ?

emptyIf all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

emptyIf all are empty where could a black pixel in S be?If all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

If all are empty where could a black pixel in S be?If all are empty where could a black pixel in S be?

if and only if







Proof: (by picture)

If all are empty then S is empty

If all are empty where could a black pixel in S be?If all are empty where could a black pixel in S be?

if and only if


(x, y)

LES(x, y)

Let LES(x, y) be the size (i.e. side length) of the largest empty square

whose bottom right is at (x, y)

Then:

If the pixel (x, y) is not empty then LES(x, y) = 0.

If (x, y) is empty and in the first row or column,

LES(x, y) = 1.

If (x, y) is empty and not in the first row or column,

LES(x, y) = min(LES(x− 1, y − 1), LES(x− 1, y), LES(x, y − 1)) + 1.




Then:



LES(x, y) = 1.



(x, y)

LES(x, y) = 0




Then:



LES(x, y) = 1.



(x, y)

LES(x, y) = 0

X




Then:



LES(x, y) = 1.



(x, y)

LES(x, y) = 1

(x, y)

X




Then:



LES(x, y) = 1.



LES(x, y) = 1

(x, y)

X




Then:



LES(x, y) = 1.



LES(x, y) = 1

(x, y)

XX




Then:



LES(x, y) = 1.



(x, y)

XX




Then:



LES(x, y) = 1.


LES(x, y) 6

(x, y)

LES(x− 1, y − 1) + 1


XX




Then:



LES(x, y) = 1.


LES(x, y) 6

(x, y)

LES(x− 1, y − 1) + 1


can’t be biggerthan this

LES(x, y)

XX




Then:



LES(x, y) = 1.


LES(x, y) 6

(x, y)

LES(x− 1, y − 1) + 1



there is a non-emptypixel in here

LES(x, y)

XX




Then:



LES(x, y) = 1.


(x, y)

LES(x, y) 6

LES(x− 1, y) + 1



LES(x, y)

XX




Then:



LES(x, y) = 1.


(x, y)

LES(x, y) 6

LES(x, y − 1) + 1



LES(x, y)

XX




Then:



LES(x, y) = 1.



(x, y)can’t be bigger

than this

LES(x, y)

XX




Then:



LES(x, y) = 1.



(x, y)

is this square always empty?


LES(x, y)

XX




Then:



LES(x, y) = 1.



(x, y)


XX




Then:



LES(x, y) = 1.



(x, y)

is this square always empty?empty

XX




Then:



LES(x, y) = 1.



(x, y)


yes, by the proof on the previous slide

empty

XX




Then:



LES(x, y) = 1.



(x, y)


yes, by the proof on the previous slide

empty

XX

X

2. Write down a recursive algorithm

We can use the recursive formula to get a recursive algorithm. . .

If pixel (x, y) is not empty

Return 0

If (x = 1) or (y = 1)

Return 1

Return min(LES(x− 1, y − 1),LES(x− 1, y),LES(x, y − 1)

)+ 1

LES(x, y)

LES(x, y) computes the size of the largest empty square


Therefore, the maximum of LES(x, y) over all x and ygives the size of the largest empty square in the whole image


We can use the recursive formula to get a recursive algorithm. . .


Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)

LES(x, y) computes the size of the largest empty square


Therefore, the maximum of LES(x, y) over all x and ygives the size of the largest empty square in the whole image

What is the time complexity of this algorithm?

How efficient is the recursive algorithm?


Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)

Let’s compute LES(4, 4). . .



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)

Let’s compute LES(4, 4). . . (and consider the recursive calls)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)

(and consider the recursive calls)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4)

(3, 3)

(3, 3) (3, 3) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4)

(3, 3)

(3, 3) (3, 3) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

computed three times :s



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

(2, 2) is computed three times just while computing this (3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

This doesn’t look good!



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

This doesn’t look good!In fact the running time of LES(n, n) is exponential in n



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

This doesn’t look good!In fact the running time of LES(n, n) is exponential in n

If T (n) is the runtime of LES(n, n)

then T (n) > 3T (n− 1)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)



Return 0

If (x = 1) or (y = 1)

Return 1


)+ 1

LES(x, y)


(4, 4)


(4, 3)(3, 4)

(2, 2) (2, 3) (3, 2) (2, 3) (2, 4)

(3, 3)

(3, 3) (3, 3) (4, 2)(3, 3)(3, 2)(3, 3)

(3, 3)

What should we do about all this repeated computation?

3. Store the solutions to subproblems


Return 0

If (x = 1) or (y = 1)

Return 1

If LES[x, y] undefined

LES[x, y]= min(MemLES(x− 1, y − 1),MemLES(x− 1, y),MemLES(x, y − 1)

)+ 1

Return LES[x, y]

MemLES(x, y)

In the MemLES version of the algorithmwe store solutions to previously computed subproblems

in an (n× n) 2D array called LES


This is called memoization (not memorization)


Return 0

If (x = 1) or (y = 1)

Return 1



)+ 1

Return LES[x, y]

MemLES(x, y)






Return 0

If (x = 1) or (y = 1)

Return 1



)+ 1

Return LES[x, y]

MemLES(x, y)



Crucially, now each entry LES[x, y] is only computed once




Return 0

If (x = 1) or (y = 1)

Return 1



)+ 1

Return LES[x, y]

MemLES(x, y)




The time complexity of computing MemLES(n, n) is now O(n2)




Return 0

If (x = 1) or (y = 1)

Return 1



)+ 1

Return LES[x, y]

MemLES(x, y)




The time complexity of computing MemLES(n, n) is now O(n2)(in fact, computing maxx,y MemLES(x, y) takes O(n2) time too

)

The dependency graph


)+ 1

LES[n, n]

What information do we need to compute LES[n, n]?

The 2D array

LES:

(for x, y > 1 and (x, y) non empty)



)+ 1

LES[n, n]


to compute

The 2D array

LES:




)+ 1

LES[n, n]


to compute

The 2D array

LES:

we need




)+ 1

LES[n, n]


to compute

The 2D array

LES:

we need

LES[n− 1, n− 1]

LES[n− 1, n]

and LES[n, n− 1]




)+ 1


The 2D array

LES:

LES[n, n− 1]to compute




)+ 1


The 2D array

LES:

LES[n, n− 1]to compute

LES[n− 1, n− 1]

and LES[n, n− 2]

LES[n− 1, n− 2]




)+ 1


The 2D array

LES:

LES[n− 1, n]to compute




)+ 1


The 2D array

LES:

we need





)+ 1


The 2D array

LES:

we need


LES[n− 2, n]

and LES[n− 1, n− 1]

LES[n− 2, n− 1]




)+ 1


The 2D array

LES:




)+ 1


The 2D array

LES: How can we use this to get an

iterative algorithm?




)+ 1


The 2D array

LES: How can we use this to get an

iterative algorithm?

Fill in the array from

the top-left!


4. Derive an iterative algorithm

For y = 1 to n

For x = 1 to n


LES[x, y] = 0

Else If (x = 1) or (y = 1)

LES[x, y] = 1

Else

LES[x, y]= min(LES[x− 1, y − 1],LES[x− 1, y],LES[x, y − 1]

)+ 1

ItLES(n)

This iterative version of the algorithm

runs in O(n2) time

and avoids making any recursive calls


For y = 1 to n

For x = 1 to n


LES[x, y] = 0

Else If (x = 1) or (y = 1)

LES[x, y] = 1

Else

LES[x, y]= min(LES[x− 1, y − 1],LES[x− 1, y],LES[x, y − 1]

)+ 1

ItLES(n)

This iterative version of the algorithm

runs in O(n2) time

and avoids making any recursive calls

Maximum of LES[x, y] over all x and ygives the size of the largest empty square in the whole image

this also takes O(n2) time

Introduction




in other words. . .

The basic idea:



3. Speed it up by storing the solutions to subproblems (memoization)(to avoid recomputing the same thing over and over)





Summary

End of part one

Part two

Weighted Interval Scheduling

Introduction




in other words. . .

The basic idea:








SummaryIntroduction


6

Problem Given an n weighted intervals,find the schedule with largest total weight

2

1

3 5 2

4

7

time

5



time

2

1

3 5 2

4

75

66



time

2

1

3 5 2

4

75

66

start time



time

2

1

3 5 2

4

75

66

start time

finish time



time

2

1

3 5 2

4

75

66

start time

finish timeweight



time

2

1

3 5 2

4

75

6

4

weight

6

start time

finish timeweight



time

2

1

3 5 2

4

75

6

4

weight

6



time

2

1

3 5 2

4

75

6

4

weight

6 4



time

2

1

3 5 2

4

75

6

4

weight

6

compatible intervals don’t overlap

4



time

2

1

3 5 2

4

75

6

4

weight

6

1

incompatible intervals overlap



time

2

1

3 5 2

4

75

6

4

weight

6

1

incompatible intervals overlap

Two intervals are compatible if they don’t overlap


6


2

1

3 5 2

4

7

time

5

4

weight



6


2

1

3 5 2

4

7

time

5

4

weight


A schedule is a set of compatible intervals



time

2

1

3 5 2

4

75

6

4

weight

6



a schedule

3 2

4


6


2

1

3 5 2

4

7

time

5

4

weight



a different schedule

2

1 7


6


2

1

3 5 2

4

7

time

5

4

weight




6


2

1

3 5 2

4

7

time

5

4

weight



The weight of a schedule is the sum ofthe weight of the intervals it contains



time

2

1

3 5 2

4

75

6

4

weight

6




a schedule

3 2

with total weight 15

4



time

2

1

3 5 2

4

75

6

4

weight




a different schedule

2

1 7




time

2

1

3 5 2

4

75

6

4

weight

6




a schedule

3 2


4



time

2

1

3 5 2

4

75

6

4

weight

6




a schedule

3 2


is this the best possible?

4



time

2

1

3 5 2

4

75

6

4

weight

6





5

6

with total weight 17a schedule

2

4



time

2

1

3 5 2

4

75

6

4

weight

6





7

a schedule


5



time

2

1

3 5 2

4

75

6

4

weight

6





7

a schedule


5

yes


6


2

1

3 5 2

4

7

time

5

4

weight


6


2

1

3 5 2

4

7

time

5

4

weight

How is the input provided?


6


2

1

3 5 2

4

7

time

5

4

weight


The intervals are given in an array A of length n


6


2

1

3 5 2

4

7

time

5

4

weight



A[i] stores a triple (si, fi, wi) which defines the i-th interval


6


2

1

3 5 2

4

7

time

5

start time, si

4

weight





6


2

1

3 5 2

4

7

time

5

start time, si

finish time, fi

4

weight





6


2

1

3 5 2

4

7

time

5

start time, si

finish time, fiweight, wi

4

weight





6


2

1

3 5 2

4

7

time

5

start time, si





4

weight, wi


6


2

1

3 5 2

4

7

time

5

start time, si





The intervals are sorted by finish time i.e. fi 6 fi+1 4

weight, wi


6


2

1

3 5 2

4

75

4

weight, wi

time


6


2

1

3 5 2

4

75

4

weight, wi

The intervals in the input are sorted by finish time

interval i finishes before interval i+ 1 finishes

time



4

weight, wi

2

1

3 5 2

4

75

62

interval 1



time



4

weight, wi

2

1

3 5 2

4

75

6

5

interval 2



time



4

weight, wi

2

1

3 5 2

4

75

6

3

interval 3



time



4

weight, wi

2

1

3 5 2

4

75

6

1

interval 4



time



4

weight, wi

2

1

3 5 2

4

75

66

interval 5



time



4

weight, wi

2

1

3 5 2

4

75

6

5

interval 6



time



4

weight, wi

2

1

3 5 2

4

75

6 4

interval 7



time



4

weight, wi

2

1

3 5 2

4

75

6

7

interval 8



time



4

weight, wi

2

1

3 5 2

4

75

6

2

interval 9



time


6


2

1

3 5 2

4

75

4

weight, wi



time

Compatible Intervals

interval p(7)

interval 7


Let p(i) be the rightmost interval (in order of finish time)which finishes before the i-th interval but doesn’t overlap it

interval p(7)

interval 7

For all i,



For all i,

interval p(4)interval 4



For all i,

interval 2



For all i,

interval 2

What is p(2)?



For all i,

interval 2

What is p(2)?

if no such interval exists, p(i) = 0



interval p(7)

interval 7

For all i,




Claim: We can precompute all p(i) in O(n logn) time

interval p(7)

interval 7

For all i,





interval p(7)

interval 7

For all i,


(and we’ll assume we did this already)




interval p(7)

interval 7

For all i,


- we’ll come back to this at the end

(and we’ll assume we did this already)


Consider some optimal scheduleO for intervals {1, 2, 3 . . . , n} with weight OPT. . .

more intervals not shown



In particular, consider the n-th interval . . .




In particular, consider the n-th interval . . .


n-th interval




n-th interval



Either the n-th interval is in scheduleO. . . or it isn’t


n-th interval




this gives us two cases to consider:


n-th interval




this gives us two cases to consider:


n-th interval

Case 2: The n-th interval is inO

Case 1: The n-th interval is not inO




n-th interval





n-th interval


is not inO




n-th interval


- scheduleO is also an optimal schedule for the problem

with the input consisting of intervals {1, 2, 3 . . . , n− 1}

is not inO



n-th interval




is not inO



so, in this case we have that OPT = OPT(n− 1)


n-th interval




is not inO



so, in this case we have that OPT = OPT(n− 1)


n-th interval


Notation: OPT(i) is the weight of an optimal schedule for intervals {1, 2, 3, . . . , i}



is not inO





n-th interval



is inO




n-th interval



is inO

The only other intervals which could be inO are {1, 2, 3, . . . p(n)}




n-th interval



is inO


not inO (they overlap)



n-th interval



is inO






n-th interval



is inO




interval p(n)



n-th interval



is inO



(the ones which don’t overlap the n-th interval)


interval p(n)



n-th interval



is inO





for the intervals {1, 2, 3 . . . , p(n)}ScheduleO with interval n removed gives an optimal schedule

interval p(n)



n-th interval



is inO






interval p(n)

(wn is the weight of interval n)



n-th interval



is inO






interval p(n)


so we have that OPT = OPT(p(n)) + wn




n-th interval



Case 2: The n-th interval is inOCase 1: The n-th interval is not inOOPT = OPT(n− 1) OPT = OPT(p(n)) + wn




n-th interval




Well, which is it?




n-th interval




Well, which is it? It’s the bigger one




n-th interval





OPT = max(OPT(n− 1),OPT(p(n)) + wn)




n-th interval





OPT = max(OPT(n− 1),OPT(p(n)) + wn)

(they both always give viable schedules)


Consider some optimal scheduleO for intervals {1, 2, 3 . . . , i} with weight OPT(i). . .


i-th interval


Case 2: The i-th interval is inOCase 1: The i-th interval is not inOOPT(i) = OPT(i− 1) OPT(i) = OPT(p(i)) + wi


OPT(i) = max(OPT(i− 1),OPT(p(i)) + wi)

(they both always give viable schedules)

(wi is the weight of interval i)


Once again, we can use the recursive formula to get a recursive algorithm. . .

If (i = 0)

Return 0

Return max(WIS(i− 1),WIS(p(i)) + wi

)WIS(i)

Therefore, WIS(n) gives the weight of the optimal schedule(for the full problem)

WIS(i) computes the weight of an optimal schedule for intervals {1, 2, 3, . . . , i}


Once again, we can use the recursive formula to get a recursive algorithm. . .

If (i = 0)

Return 0


)WIS(i)

Therefore, WIS(n) gives the weight of the optimal schedule(for the full problem)

What is the time complexity of this algorithm?

WIS(i) computes the weight of an optimal schedule for intervals {1, 2, 3, . . . , i}


consider this simple input with n = 6

11

11

11

If (i = 0)

Return 0


)WIS(i)



If (i = 0)

Return 0


)WIS(i)

11

11

11

(the best schedule has weight 3)



If (i = 0)

Return 0


)WIS(i)

11

11

11


further, for all i, p(i) = i− 2



If (i = 0)

Return 0


)WIS(i)

11

11

11


further, for all i, p(i) = i− 2

so WIS(i) makes recursive calls to WIS(i− 1) and WIS(i− 2)


If (i = 0)

Return 0


)WIS(i)



If (i = 0)

Return 0


)WIS(i)


6WIS(6)


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 4


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 23

334


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 2

02 1 1

33

2 1

343 23


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 2

02

1 0

1

0 0

1

0 0

33

2

1 0

1

0 0

343 2

02 1 1

3


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 2

02

1 00 0

1

0 0

1

0 0

33

2

1 00 0

1

0 0

343 2

02

1 0

1

0 0

1

0 0

3


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 2

02

1 00 0

1

0 0

1

0 0

33

2

1 00 0

1

0 0

343 2

02

1 00 0

1

0 0

1

0 0

3


If (i = 0)

Return 0


)WIS(i)


6WIS(6)

5 43 2

02

1 00 0

1

0 0

1

0 0

33

2

1 00 0

1

0 0

343 2

02

1 00 0

1

0 0

1

0 0

3

This doesn’t look good (but it does look familiar)


11

11

11

If (i = 0)

Return 0


)WIS(i)

if we extend this input in the same way. . .


11

11

11

If (i = 0)

Return 0


)WIS(i)


1


11

11

11

If (i = 0)

Return 0


)WIS(i)


11


11

11

11

If (i = 0)

Return 0


)WIS(i)


11

1


11

11

11

If (i = 0)

Return 0


)WIS(i)


11

11


11

11

11

If (i = 0)

Return 0


)WIS(i)


11

11

Given n intervals set out in this manner,


11

11

11

If (i = 0)

Return 0


)WIS(i)


11

11


WIS(n) runs in exponential time


11

11

11

If (i = 0)

Return 0


)WIS(i)


11

11


WIS(n) runs in exponential time

If T (n) is the run time of WIS(n) using these intervals

then T (n) > 2T (n− 2)


If (i = 0)

Return 0

If WIS[i] undefined

WIS[i]= max(MemWIS(i− 1),MemWIS(p(i)) + wi

)Return WIS[i]

MemWIS(i)


In the MemWIS version of the algorithmwe store solutions to previously computed subproblems

in an n length array called WIS

If (i = 0)

Return 0

If WIS[i] undefined


)Return WIS[i]

MemWIS(i)


(we have memoized the algorithm)



If (i = 0)

Return 0

If WIS[i] undefined


)Return WIS[i]

MemWIS(i)





Each entry WIS[i] is only computed once

If (i = 0)

Return 0

If WIS[i] undefined


)Return WIS[i]

MemWIS(i)






The time complexity of computing MemWIS(n) is now O(n)

If (i = 0)

Return 0

If WIS[i] undefined


)Return WIS[i]

MemWIS(i)






The time complexity of computing MemWIS(n) is now O(n)

If (i = 0)

Return 0

If WIS[i] undefined


)Return WIS[i]

MemWIS(i)

because every recursion causes an unfilled entry to be filled in the array


What information do we need to compute WIS[i]?

The array

WIS:



The array

WIS:

WIS[i]



to compute

The array

WIS:

WIS[i− 1] and WIS[p(i)]

WIS[i]

WIS[i] we need



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]

both of which are to the left of WIS[i]



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]

both of which are to the left of WIS[i](somewhere)



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]


all of the dependencies go left. . .



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]



This suggests another

iterative algorithm



to compute

The array

WIS:


WIS[i]

WIS[i] we need

WIS[i− 1]

WIS[p(i)]



This suggests another

iterative algorithm

Fill in the array from

the left again


This is an iterative dynamic programming algorithm

it runs in O(n) time

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n

WIS[i]= max(WIS[i− 1],WIS[p(i)]+wi

)Return WIS[i]

for Weighted Interval Scheduling


This is an iterative dynamic programming algorithm

it runs in O(n) time

. . . but it requires than you precomputed all the p(i) values

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

for Weighted Interval Scheduling

How do you find all those p(i) values?

Revised Claim: We can precompute any p(i) in O(logn) time

interval p(7)

interval 7



interval p(7)

interval 7

Recall that si is the start time of interval iand fi is the finish time of interval i



interval p(7)

interval 7

Recall that si is the start time of interval i

We want to find the unique value j = p(i) such that

and fi is the finish time of interval i

fj < si < fj+1.



interval p(7)

interval 7




fj < si < fj+1.

si

f1 f2 f3 f4 f5f6f7



interval p(7)

interval 7




fj < si < fj+1.

As the input is sorted by finish times, we can find j by binary search in O(logn) time

si

f1 f2 f3 f4 f5f6f7



interval p(7)

interval 7



interval p(7)

interval 7

Original Claim: We can precompute all p(i) in O(n logn) time



interval p(7)

interval 7

Original Claim: We can precompute all p(i) in O(n logn) time

(by using the revised claim n times)

Wait, did you want the actual schedule?

but doesn’t find the actual scheduleItWIS(n) finds the weight of the optimal schedule

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]


but doesn’t find the actual schedule

There is an optimal schedule for {1, 2, . . . , i} containing

interval i if and only if

WIS[i− 1]6WIS[p(i)]+wi

ItWIS(n) finds the weight of the optimal schedule

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]


but doesn’t find the actual schedule




(by the argument we saw earlier)

ItWIS(n) finds the weight of the optimal schedule

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]





(by the argument we saw earlier)

ItWIS(n) finds the weight of the optimal scheduleand FINDWIS(n) finds the actual schedule

FindWIS(i)

If (i = 0)

Return nothing

If WIS[i− 1]6WIS[p(i)]+wi

Return FindWIS(p(i)) then i

Return FindWIS(i− 1)

This is called backtracking and works for lots of Dynamic Programming algorithms

ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

The final algorithm


The final algorithm:

Step 1: Find all the p(i) values

Step 2: Run ItWIS(n) to find the optimal weight

Step 3: Run FindWIS(n) to find the schedule

FindWIS(i)

If (i = 0)

Return nothing




ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

The final algorithm






O(n logn) time

FindWIS(i)

If (i = 0)

Return nothing




ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

The final algorithm






O(n logn) time

O(n) time

FindWIS(i)

If (i = 0)

Return nothing




ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

The final algorithm






O(n logn) time

O(n) time

O(n) time

FindWIS(i)

If (i = 0)

Return nothing




ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

The final algorithm






O(n logn) time

O(n) time

O(n) time

Overall this takes O(n logn) time

FindWIS(i)

If (i = 0)

Return nothing




ItWIS(n)

If (i = 0)

Return 0

For i = 1 to n


)Return WIS[i]

Introduction




in other words. . .

The basic idea:








SummaryIntroduction

Summary

Dynamic Programming

Education

Transcript of Dynamic Programming