Lecture 18.

Dynamic Programming

Introduction to Algorithms
Da Nang University of Science and Technology

Dang Thien Binh

[email protected]

Intelligent Networking Laboratory

Copyright 2000-2022 Intelligent
DANG
Networking
THIEN BINH
Laboratory
1/32
 Dynamic Programming
 Dynamic programming solves optimization problems by combining
solutions to subproblems
 “Programming” refers to a tabular method with a series of choices,
not “coding”
 Recall the divide-and-conquer approach
 Partition the problem into independent subproblems
 Solve the subproblems recursively
 Combine solutions of subproblems
 Dynamic programming is applicable when subproblems are not
independent, i.e., subproblems share subsubproblems
 Solve every subsubproblem only once and store the answer for use
when it reappears
 A divide-and-conquer approach will do more work than necessary

Intelligent Networking Laboratory DANG THIEN BINH 2/32

 Dynamic Programming Solution
 4 Steps
 1. Characterize the structure of an optimal solution
 2. Recursively define the value of an optimal solution
 3. Compute the value of an optimal solution in a bottom-
up fashion
 4. Construct an optimal solution from computed
information

Intelligent Networking Laboratory DANG THIEN BINH 3/32

 Matrix Multiplication (1/3)
 Matrix multiplication
 Two matrices 𝐴 and 𝐵 can be multiplied
 The number of columns of 𝐴 must equal the number of
rows of 𝐵
 𝐴(𝑝𝑞) x 𝐵(𝑞𝑟) → 𝐶(𝑝𝑟)
 The number of scalar multiplications is 𝑝𝑞𝑟
 For a 𝑝𝑞 matrix 𝐴 and a 𝑞𝑟 matrix 𝐵, the product 𝐴𝐵 is a
𝑝𝑟 matrix 𝐶
𝑞

𝐴𝐵 = 𝐶 = [𝑐𝑖,𝑗 ]𝑝×𝑟 ≡ ෍ 𝑎𝑖,𝑘 𝑏𝑘,𝑗

𝑘=1 𝑝×𝑟

Intelligent Networking Laboratory DANG THIEN BINH 4/32

 Matrix Multiplication (2/3)

Intelligent Networking Laboratory DANG THIEN BINH 5/32

 Matrix Multiplication (3/3)
 Matrix multiplication

Intelligent Networking Laboratory DANG THIEN BINH 6/32

 Matrix-Chain Multiplication (1/2)
 Matrix multiplication
 𝐴(𝑝𝑞) x 𝐵(𝑞𝑟) → 𝐶(𝑝𝑟)
 The number of scalar multiplications is 𝑝𝑞𝑟
 e.g. < 𝐴1(10100), 𝐴2(1005), 𝐴3(550) >
 ((𝐴1𝐴2)𝐴3) → 10x100x5 + 10x5x50 = 5000 + 2500 = 7,500
 (𝐴1(𝐴2𝐴3)) → 100x5x50 + 10x100x50 = 25000 + 50000 = 75,000
 10 times faster
 Given a sequence (chain) <𝐴1, 𝐴2,…, 𝐴𝑛> of n matrices to be
multiplied, where 𝑖 = 1,2, … , 𝑛, matrix 𝐴𝑖 has dimension 𝑝𝑖−1 𝑝𝑖,
fully parenthesize the product 𝐴1𝐴2 … 𝐴𝑛 in a way that
minimizes the number of scalar multiplications
 Determine an order for multiplying matrices that has the lowest
cost

Intelligent Networking Laboratory DANG THIEN BINH 7/32

 Matrix-Chain Multiplication (2/2)
 Determine an order for multiplying matrices
that has the lowest cost
 Counting the number of parenthesizations
 𝐴1 𝐴2 … 𝐴𝑘 𝐴𝑘+1 … 𝐴𝑛−1 𝐴𝑛
1 𝑖𝑓 𝑛 = 1
 𝑃(𝑛) =
𝑛−1
( 𝟐𝒏 )
෍ 𝑃(𝑘)𝑃(𝑛 − 𝑘) 𝑖𝑓 𝑛 ≥ 2
𝑘=1

 Exercise 15.2-3 on page 338

 Impractical to check all possible parenthesizations

Intelligent Networking Laboratory DANG THIEN BINH 8/32

 Step 1 (1/3)
 The structure of an optimal parenthesization
 Notation: 𝐴𝑖..𝑗
 Result from evaluating 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 (𝑖 < 𝑗)
 Any parenthesization of 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 must split
the product between 𝐴𝑘 and 𝐴𝑘+1 for some integer 𝑘
in the range 𝑖 ≤ 𝑘 < 𝑗
 The cost of this parenthesization
 cost of computing 𝐴𝑖..𝑘 + cost of computing 𝐴𝑘+1..𝑗
+ cost of multiplying 𝐴𝑖..𝑘 and 𝐴𝑘+1..𝑗 together

Intelligent Networking Laboratory DANG THIEN BINH 9/32

 Step 1 (2/3)
 Suppose that an optimal parenthesization of
𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 splits the product between 𝐴𝑘 and 𝐴𝑘+1
 The parenthesization of the prefix sub-chain 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑘
must be an optimal parenthesization of 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗
 The parenthesization of the postfix sub-chain 𝐴𝑘+1 𝐴𝑖+1 … 𝐴𝑗
must be an optimal parenthesization of 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗

 That is, the optimal solution to the problem contains

within it the optimal solution to subproblems

Intelligent Networking Laboratory DANG THIEN BINH 10/32

 Step 1 (3/3)
Minimal
Cost_A1..6 + Cost_A7..9+p0p6p9
A1A2A3A4A5A6A7A8A9

Suppose ((A1A2)(A3((A4A5)A6))) ((A7A8)A9) is optimal

Then (A1A2) (A3((A4A5)A6)) must be optimal for A1A2A3A4A5A6

Otherwise, if (A1(A2 A3)) ((A4A5)A6) is optimal for A1A2A3A4A5A6

Then ((A1(A2 A3)) ((A4A5)A6)) ((A7A8)A9) will be better than

((A1A2)(A3((A4A5)A6))) ((A7A8)A9)
Contradiction!
Intelligent Networking Laboratory DANG THIEN BINH 11/32
 Step 2
 A Recursive Solution
 Subproblem
 Determine the minimum cost of a parenthesization of
𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 (1 ≤ 𝑖 ≤ 𝑗 ≤ 𝑛)
 𝑚[𝑖, 𝑗]: the minimum number of scalar multiplications
needed to compute the matrix 𝐴𝑖..𝑗
 𝑚[𝑖, 𝑗] = 𝑚[𝑖, 𝑘] + 𝑚[𝑘 + 1, 𝑗] + 𝑝𝑖−1 𝑝𝑘 𝑝𝑗
 However, we do not know the value of 𝑘 (𝑠[𝑖, 𝑗]),
so we have to try all j-i possibilities
0 if i = j 
 
m[i, j ] =  {m[i, k ] + m[k + 1, j ]} + pi −1 pk p j if i  j 
min
 ik  j 

 A recursive solution takes exponential time

Intelligent Networking Laboratory DANG THIEN BINH 12/32

 Step 3 (1/3)
 Computing the optimal costs
 How much subproblems in total?
 One for each choice of 𝑖 and 𝑗 satisfying 1 ≤ 𝑖 ≤ 𝑗 ≤ 𝑛
 (𝑛2)
 MATRIX-CHAIN-ORDER(p)
 Input: a sequence 𝑝 = < 𝑝0, 𝑝1, 𝑝2, … , 𝑝𝑛> (length[𝑝] = 𝑛 + 1)
 Try to fill in the table 𝑚 in a manner that corresponds to
solving the parenthesization problem on matrix chains of
increasing length
 Lines 4-12: compute 𝑚[𝑖, 𝑖 + 1], 𝑚[𝑖, 𝑖 + 2], … each time

Intelligent Networking Laboratory DANG THIEN BINH 13/32

 Step 3 (2/3)
 e.g. < 𝐴1(83), 𝐴2(35), 𝐴3(510) >
 𝑚12 = 8x3x5 = 120
 𝑚23 = 3x5x10 = 150
 𝑚13 = 390
 𝑚11 + 𝑚23 + 𝑝0𝑝1𝑝3 = 0 + 150 + 8x3x10 = 390
 𝑚12 + 𝑚33 + 𝑝0𝑝2𝑝3 = 120 + 0 + 8x5x10 = 520
j
i j
1 2 3
2 3 i
m 0 120 390 1
s 1 1 1
0 150 2
2 2
0 3

Intelligent Networking Laboratory DANG THIEN BINH 14/32

 Step 3 (3/3)

𝑶(𝒏𝟑),  (𝒏𝟑) ⇒ (𝒏𝟑) running time

(𝒏𝟐) space

Intelligent Networking Laboratory DANG THIEN BINH 15/32

 m[3,4]+m[5,5] + 15x10x20 = 750 + 0 + 3000 = 3750
m[3,5] = min
m[3,3]+m[4,5] + 15x5x20 = 0 + 1000 + 1500 = 2500

l =3
l=2

35x15x5=2625 10x20x25=5000

Intelligent Networking Laboratory DANG THIEN BINH 16/32

 Step 4 (1/2)
 Constructing an optimal solution
 Each entry 𝑠[𝑖, 𝑗] records the value of k such that
the optimal parenthesization of 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 splits the
product between 𝐴𝑘 and 𝐴𝑘+1
 𝐴1..𝑛 ⇒ 𝐴1..𝑠[1..𝑛] 𝐴𝑠 1..𝑛 +1..𝑛
 𝐴1..𝑠[1..𝑛] ⇒ 𝐴1..𝑠[1, 𝑠[1..𝑛]] 𝐴𝑠 1, 𝑠[1..𝑛 ]+1..𝑠[1..𝑛]
 Recursive…


Intelligent Networking Laboratory DANG THIEN BINH 17/32

 Step 4 (2/2)
 Constructing an optimal solution


 (( 𝐴1 ( 𝐴2 𝐴3 ) ) ( ( 𝐴4 𝐴5 ) 𝐴6 ))

Intelligent Networking Laboratory DANG THIEN BINH 18/32

 Elements of Dynamic Programming (1/2)
 Optimal substructure
 If an optimal solution contains within it optimal solutions
to subproblems
 Build an optimal solution from optimal solutions to
subproblems
 Example
 Matrix-chain multiplication
 An optimal parenthesization of 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑗 that splits the
product between 𝐴𝑘 and 𝐴𝑘+1 contains within it optimal
solutions to the problem of parenthesizing 𝐴𝑖 𝐴𝑖+1 … 𝐴𝑘 and
𝐴𝑘+1 𝐴𝑘+2 … 𝐴𝑗

Intelligent Networking Laboratory DANG THIEN BINH 19/32

 Elements of Dynamic Programming (2/2)
 Overlapping subproblems
 The space of subproblems must be small in the sense that a
recursive algorithm for the problem solves the same
subproblems over and over, rather than always generating new
subproblems
 Typically, the total number of distinct subproblems is a
polynomial in the input size
 Divide-and-Conquer is suitable usually generate brand-
new problems at each step of the recursion

Intelligent Networking Laboratory DANG THIEN BINH 20/32

 Characteristics of Optimal Substructure
 How many subproblems are used in an optimal
solution to the original problem?
 Matrix-Chain scheduling
 2 (𝐴1𝐴2 … 𝐴𝑘 and 𝐴𝑘+1 𝐴𝑘+2 … 𝐴𝑗 )

 How many choices we have in determining which

subproblems to use in an optimal solution?
 Matrix-chain scheduling: 𝑗 − 𝑖 (choice for 𝑘)
 Informally, the running time of a dynamic-
programming algorithm 𝑠 on: the number of
subproblems overall and how many choices we look
at for each subproblem
 Matrix-Chain scheduling: (𝑛2) ∗ 𝑂(𝑛) = 𝑂(𝑛3)

Intelligent Networking Laboratory DANG THIEN BINH 21/32

 Overlapping Subproblems (1/2)

Intelligent Networking Laboratory DANG THIEN BINH 22/32

 Overlapping Subproblems (2/2)

m[3,4] is computed twice

Intelligent Networking Laboratory DANG THIEN BINH 23/32

 Recursive Procedure for
Matrix-Chain Multiplication
 The time to compute 𝑚[1, 𝑛] is at least exponential
in 𝑛
 𝑇(1) ≥ 1
𝑛−1

𝑇(𝑛) ≥ 1 + ෍ (𝑇(𝑘) + 𝑇(𝑛 − 𝑘) + 1)

𝑘=1
𝑛−1

𝑇(𝑛) ≥ 2 ෍ 𝑇(𝑖) + 𝑛
𝑖=1

 Prove 𝑇(𝑛) = 𝛺(2𝑛) using the substitution method

 Show that 𝑇(𝑛) ≥ 2𝑛−1
𝑛−1 𝑛−2

 𝑇(𝑛) ≥ 2 ෍ 2𝑖−1 + 𝑛 = 2 ෍ 2𝑖 + 𝑛
𝑖=1 𝑖=0
= 2(2𝑛−1 − 1) + 𝑛 = (2𝑛 − 2) + 𝑛 ≥ 2𝑛−1

Intelligent Networking Laboratory DANG THIEN BINH 24/32

 Memoization
 A variation of dynamic programming that often
offers the efficiency of the usual dynamic programming
approach while maintaining a top-down strategy
 Memoize the natural, but inefficient, recursive algorithm
 Maintain a table with subproblem solutions, but
the control structure for filling in the table is more like
the recursive algorithm
 Memoization for matrix-chain multiplication
 Calls in which 𝑚 𝑖, 𝑗 =  ⇒ (𝑛2) calls
 Calls in which 𝑚 𝑖, 𝑗 <  ⇒ 𝑂(𝑛3) calls
 Turns an (2𝑛)-time into an 𝑂(𝑛3)-time algorithm

Intelligent Networking Laboratory DANG THIEN BINH 25/32

 Memoization

Intelligent Networking Laboratory DANG THIEN BINH 26/32

 Lookup-Chain(p, i, j)

Intelligent Networking Laboratory DANG THIEN BINH 27/32

 Lookup-Chain(p, i, j)

Intelligent Networking Laboratory DANG THIEN BINH 28/32

 Dynamic Programming vs. Memoization
 If all subproblems must be solved at least once, a
bottom-up dynamic-programming algorithm usually
outperforms a top-down memoized algorithm by a
constant factor
 No overhead for recursion and less overhead for
maintaining table
 There are some problems for which the regular pattern of
table accesses in the dynamic programming algorithm can
be exploited to reduce the time or space requirements
even further

Intelligent Networking Laboratory DANG THIEN BINH 29/32

 Self-Study
 Two more dynamic-programming problems
 Section 15.4 Longest Common Subsequence
 Section 15.5 Optimal Binary Search Trees

Intelligent Networking Laboratory DANG THIEN BINH 30/32

 Longest Common Subsequence (LCS)
 Problem: Given two sequences 𝑋=<𝑥1, … , 𝑥𝑚> and
𝑌=<𝑦1, … , 𝑦𝑛>, find the longest subsequence
𝑍=<𝑧1, … , 𝑧𝑘 > that is common to 𝑋 and 𝑌
 A subsequence is a subset of elements from the sequence with
strictly increasing order (not necessarily contiguous)
 There are 2𝑚 subsequences of 𝑋 → checking all subsequences is
impractical for long sequences
 Example: 𝑋=<𝐴, 𝐵, 𝐶, 𝐵, 𝐷, 𝐴, 𝐵> and 𝑌=<𝐵, 𝐷, 𝐶, 𝐴, 𝐵, 𝐴>
 Common subsequences: <𝐴>; <𝐵>; <𝐶>; <𝐷>; <𝐴, 𝐴>; <𝐵, 𝐵>;
<𝐵, 𝐶, 𝐴>; <𝐵, 𝐶, 𝐵>; <𝐶, 𝐵, 𝐴>; etc.
 The longest common subsequences: <𝐵, 𝐶, 𝐵, 𝐴>; <𝐵, 𝐷, 𝐴, 𝐵>

Intelligent Networking Laboratory DANG THIEN BINH 31/32

 Step 1: Optimal Structure of an LCS
 Let 𝑋=<𝑥1, … , 𝑥𝑚> and 𝑌=<𝑦1, … , 𝑦𝑛> be sequences, and
let 𝑍=<𝑧1, … , 𝑧𝑘 > be any LCS of 𝑋 and 𝑌
 If 𝑥𝑚 = 𝑦𝑛, then 𝑧𝑘 = 𝑥𝑚 = 𝑦𝑛 and 𝑍𝑘−1 is an LCS of 𝑋𝑚−1 and 𝑌𝑛−1

 If 𝑥𝑚 ≠ 𝑦𝑛, then 𝑧𝑘 ≠ 𝑥𝑚 implies that 𝑍 is an LCS of 𝑋𝑚−1 and 𝑌

 If 𝑥𝑚 ≠ 𝑦𝑛, then 𝑧𝑘 ≠ 𝑦𝑛 implies that 𝑍 is an LCS of 𝑋 and 𝑌𝑛−1

Intelligent Networking Laboratory DANG THIEN BINH 32/32

 Step 2: Recursive Solution (1/2)
 Overlapping-subproblems

Intelligent Networking Laboratory DANG THIEN BINH 33/32

 Step 2: Recursive Solution (2/2)
 Define 𝑐[𝑖, 𝑗] = length of LCS for 𝑋𝑖 and 𝑌𝑗

Intelligent Networking Laboratory DANG THIEN BINH 34/32

 Step 3: Computing the length of an LCS

𝑏[𝑖, 𝑗] points to the table entry

corresponding to the optimal
subproblem solution chosen
when computing 𝑐[𝑖, 𝑗]

LCS-LENGTH(X, Y) is (𝒎𝒏)

Intelligent Networking Laboratory DANG THIEN BINH 35/32

 Step 4: Constructing an LCS

PRINT-LCS(b, X, i, j) is 𝑶(𝒎 + 𝒏)

Intelligent Networking Laboratory DANG THIEN BINH 36/32

 Thanks to contributors
Mr. Phuoc-Nguyen Bui (2022)
Dr. Thien-Binh Dang (2017 – 2022)
Prof. Hyunseung Choo (2001 – 2022)

Intelligent Networking Laboratory

Copyright 2000-2022 Intelligent
DANGNetworking
THIEN BINH
Laboratory
37/32