Module-4 : Dynamic Programming

Contents

1. Introduction to Dynamic Programming 4. Optimal Binary Search Trees

General method with Examples 5. Knapsack problem
2. Transitive Closure: 6. Bellman-Ford Algorithm
Warshall’s Algorithm, 7. Travelling Sales Person problem
3. All Pairs Shortest Paths:
Floyd's Algorithm,
 2

1. Introduction to Dynamic Programming

Dynamic programming is a technique for solving problems with overlapping subproblems.
Typically, these subproblems arise from a recurrence relating a given problem’s solution to
solutions of its smaller subproblems. Rather than solving overlapping subproblems again and
again, dynamic programming suggests solving each of the smaller subproblems only once
and recording the results in a table from which a solution to the original problem can then be
obtained. [From T1]
The Dynamic programming can be used when the solution to a problem can be viewed as the
result of sequence of decisions. [ From T2]. Here are some examples.

Example 1

Example 2

Example 3

Example 4
3
4
 5

2. Warshall’s Algorithm,
Definition: The transitive closure of a directed graph with n vertices can be defined as the n
× n boolean matrix T = {tij}, in which the element in the ith row and the jth column is 1 if there
exists a nontrivial path (i.e., directed path of a positive length) from the i th vertex to the jth
vertex; otherwise, tij is 0.
Example: An example of a digraph, its adjacency matrix, and its transitive closure is given
below.

(a) Digraph. (b) Its adjacency matrix. (c) Its transitive closure.

We can generate the transitive closure of a digraph with the help of depth first search or
breadth-first search. Performing either traversal starting at the i th vertex gives the information
about the vertices reachable from it and hence the columns that contain 1’s in the ith row of
the transitive closure. Thus, doing such a traversal for every vertex as a starting point yields
the transitive closure in its entirety.
Since this method traverses the same digraph several times, we can use a better algorithm
called Warshall’s algorithm. Warshall’s algorithm constructs the transitive closure through
a series of n × n boolean matrices:

Each of these matrices provides certain information about directed paths in the digraph.
Specifically, the element𝑟𝑖𝑗(𝑘) in the ith row and jth column of matrix R(k) (i, j = 1, 2, . . . , n, k =
0, 1, . . . , n) is equal to 1 if and only if there exists a directed path of a positive length from
the ith vertex to the jth vertex with each intermediate vertex, if any, numbered not higher than
k.
Thus, the series starts with R(0), which does not allow any intermediate vertices in its paths;
hence, R(0) is nothing other than the adjacency matrix of the digraph. R(1) contains the
information about paths that can use the first vertex as intermediate. The last matrix in the
series, R(n),reflects paths that can use all n vertices of the digraph as intermediate and hence is
nothing other than the digraph’s transitive closure.
This means that there exists a path from the ith vertex vi to the jth vertex vj with each
intermediate vertex numbered not higher than k:
vi, a list of intermediate vertices each numbered not higher than k, vj . --- (*)
Two situations regarding this path are possible.
 6

1. In the first, the list of its intermediate vertices does not contain the kth vertex. Then this
path from vi to vj has intermediate vertices numbered not higher than k−1. i.e. rij(k−1) = 1
th
2. The second possibility is that path (*) does contain the k vertex vk among the
intermediate vertices. Then path (*) can be rewritten as;
vi, vertices numbered ≤ k − 1, vk, vertices numbered ≤ k − 1, vj .
i.e r(k−1) = 1 and r(k−1) = 1
ik kj

Thus, we have the following formula for generating the elements of matrix R(k) from the
elements of matrix R(k−1)

The Warshall’s algorithm works based on the above formula.

As an example, the application of Warshall’s algorithm to the digraph is shown below. New
1’s are in bold.
 7

Analysis
Its time efficiency is Θ(n3). We can make the algorithm to run faster by treating matrix rows
as bit strings and employ the bitwise or operation available in most modern computer
languages.
Space efficiency: Although separate matrices for recording intermediate results of the
algorithm are used, that can be avoided.

3. All Pairs Shortest Paths using Floyd's Algorithm,

Problem definition: Given a weighted connected graph (undirected or directed), the all-pairs
shortest paths problem asks to find the distances—i.e., the lengths of the shortest paths - from
each vertex to all other vertices.
Applications: Solution to this problem finds applications in communications, transportation
networks, and operations research. Among recent applications of the all-pairs shortest-path
problem is pre-computing distances for motion planning in computer games.
We store the lengths of shortest paths in an n x n matrix D called the distance matrix: the
element dij in the ith row and the jth column of this matrix indicates the length of the shortest
path from the ith vertex to the jth vertex.

(a) Digraph. (b) Its weight matrix. (c) Its distance matrix
We can generate the distance matrix with an algorithm that is very similar to Warshall’s
algorithm. It is called Floyd’s algorithm.
Floyd’s algorithm computes the distance matrix of a weighted graph with n vertices through a
series of n × n matrices:
 8

th th (k)
The element 𝑑(𝑘)
𝑖𝑗 in the i row and the j column of matrix D (i, j = 1, 2, . . . , n, k = 0, 1, .
. . , n) is equal to the length of the shortest path among all paths from the i th vertex to the jth
vertex with each intermediate vertex, if any, numbered not higher than k.
As in Warshall’s algorithm, we can compute all the elements of each matrix D(k) from its
immediate predecessor D(k−1)

If 𝑑𝑖𝑗
(𝑘) = 1, then it means that there is a path;

vi, a list of intermediate vertices each numbered not higher than k, vj .

We can partition all such paths into two disjoint subsets: those that do not use thek th vertex vk
as intermediate and those that do.
i. Since the paths of the first subset have their intermediate vertices numbered not higher
than k − 1, the shortest of them is, by definition of our matrices, of length 𝑑𝑖𝑗
(𝑘−1)

ii. In the second subset the paths are of the form

vi, vertices numbered ≤ k − 1, vk, vertices numbered ≤ k − 1, vj .

The situation is depicted symbolically in Figure, which shows

the underlying idea of Floyd’s algorithm.

Taking into account the lengths of the shortest paths in both subsets leads to the following
recurrence:

Analysis: Its time efficiency is Θ(n3), similar to the warshall’s algorithm.

 9

Application of Floyd’s algorithm to the digraph is shown below. Updated elements are shown
in bold.

4. Optimal Binary Search Trees

A binary search tree is one of the most important data structures in computer science. One of
its principal applications is to implement a dictionary, a set of elements with the operations of
searching, insertion, and deletion.
The binary search tree for which the average number of comparisons in a search is the
smallest as possible is called as optimal binary search tree. If probabilities of searching for
elements of a set are known an optimal binary search tree can be constructed.
As an example, consider four keys A, B, C,
and D to be searched for with probabilities
0.1, 0.2, 0.4, and 0.3, respectively. The figure
depicts two out of14 possible binary search
trees containing these keys.
 10

The average number of comparisons in a successful search in the first of these trees is
0.1 * 1+ 0.2 * 2 + 0.4 *3+ 0.3 * 4 = 2.9, and for the second one it is 0.1 * 2 + 0.2 * 1+ 0.4 * 2
+ 0.3 * 3= 2.1. Neither of these two trees is, in fact, optimal.
For our tiny example, we could find the optimal tree by generating all 14binary search trees
with these keys. As a general algorithm, this exhaustive-search approach is unrealistic: the
total number of binary search trees with n keys is equal to the nth Catalan number,

which grows to infinity as fast as 4n/ n1.5

Problem definition: Given a sorted array a1, . . . , an of search keys and an array p1, . . . ,
pn of probabilities of searching, where pi is the probability of searches to ai. Construct a
binary search tree of all keys such that, smallest average number of comparisons made in a
successful search.
Solution using Dynamic programming
So let a1, . . . ,an be distinct keys ordered from the smallest to the largest and let p 1, . . . , pn be
the probabilities of searching for them.
Let C(i, j) be the smallest average number of comparisons made in a successful search in a
binary search treeTijmade up of keys ai, . . , aj, where i, j are some integer indices, 1≤ i ≤ j ≤ n.
We are interested just in C(1, n). Following the classic dynamic programming approach, we
will find values of C(i, j) for all smaller instances of the problem.
To derive a recurrence underlying a dynamic programming algorithm, wewill consider all
possible ways to choose a root ak among the keys ai, . . . ,aj . Forsuch a binary search tree
(Figure 8.8), the root contains key ak, the left subtreeTik−1contains keys ai, . . . , ak−1 optimally
arranged, and the right subtree Tjk+1contains keys ak+1, . . . , aj also optimally arranged. Here
we are taking advantage of the principle of optimality.
 11

If we count tree levels starting with 1 to make the comparison numbers equal the keys’ levels,
the following recurrence relation is obtained:
 12

The two-dimensional table in Figure 8.9 shows the values needed for computing C(i, j) by
formula (8.8): they are in row i and the columns to the left of column j and in column j and
the rows below row i. The arrows point to the pairs of entries whose sums are computed in
order to find the smallest one to be recorded as the value of C(i, j). This suggests filling the
table along its diagonals, starting with all zeros on the main diagonal and given probabilities
pi, 1≤ i ≤ n, right above it and moving toward the upper right corner.
The algorithm we just sketched computes C(1, n)—the average number of comparisons for
successful searches in the optimal binary tree. If we also want to get the optimal tree itself,
we need to maintain another two-dimensional table to record the value of k for which the
minimum in (8.8) is achieved. The table has the same shape as the table in Figure 8.9 and is
filled in the same manner, starting with entries R(i, i) = i for 1≤ i ≤ n. When the table is filled,
its entries indicate indices of the roots of the optimal subtrees, which makes it possible to
reconstruct an optimal tree for the entire set given.
Example: Let us illustrate the algorithm by applying it to the four-key set we used at the
beginning of this section:

The initial tables look like this:

Let us compute C(1, 2):

Thus, out of two possible binary trees containing the first two keys, A and B, the root of the
optimal tree has index 2 (i.e., it contains B), and the average number of comparisons in a
successful search in this tree is 0.4. On finishing the computations we get the following final
tables:
 13

Thus, the average number of key comparisons in the optimal tree is equal to1.7. Since R(1, 4)
= 3, the root of the optimal tree contains the third key, i.e., C. Its left subtree is made up of
keys A and B, and its right subtree contains just key D. To find the specific structure of these
subtrees, we find first their roots by consulting the root table again as follows. Since R(1, 2) =
2, the root of the optimal tree containing A and B is B, with A
being its left child (and the root of the one-node tree: R(1, 1) =
1). Since R(4, 4) = 4, the root of this one-node optimal tree
consist only key D. Figure given below presents the optimal
tree in its entirety.
Here is Pseudocode of the dynamic programming algorithm.

5. Knapsack problem
We start this section with designing a dynamic programming algorithm for the knapsack
problem: given n items of known weights w1, . . . ,wn and valuesv1, . . . , vn and a knapsack of
capacity W, find the most valuable subset of the items that fit into the knapsack. To design a
dynamic programming algorithm, we need to derive a recurrence relation that expresses a
solution to an instance of the knapsack problem in terms of solutions to its smaller sub
instances. Let us consider an instance defined by the first i items, 1≤ i ≤ n, with weights w1, . .
. ,wi, values v1, . . . , vi , and knapsack capacity j, 1 ≤ j ≤ W. Let F(i, j) be the value of an
optimal solution to this instance. We can divide all the subsets of the first i items that fit the
knapsack of capacity j into two categories: those that do not include the i th item and those that
do. Note the following:
 14

i. Among the subsets that do not include the ith item, the value of an optimal subset is,
by definition, i.e F(i , j) = F(i − 1, j).
ii. Among the subsets that do include the ith item (hence, j − wi≥ 0), an optimal subset is
made up of this item and an optimal subset of the first i−1 items that fits into the
knapsack of capacity j − wi. The value of such an optimal subset is vi+ F(i − 1, j − wi).
Thus, the value of an optimal solution among all feasible subsets of the first I items is the
maximum of these two values.

It is convenient to define the initial conditions as follows:

F(0, j) = 0 for j ≥ 0 and F(i, 0) = 0 for i ≥ 0.
Our goal is to find F(n, W), the maximal value of a subset of the n given items that fit into
the knapsack of capacity W, and an optimal subset itself.

The algorithm for the knapsack problem can be stated as follows

Input: n – total items, W – capacity of the knapsack
wi– weight of the ith item, vi– value of the ith item,
Output: F(i, j) be the value of an optimal solution to this instance considering first i items
with capacity j. F(n,W) is the optimal solution
Method:
 15

Example-1:Let us consider the instance given by the following data:

The dynamic programming table, filled by applying formulas is given below

Thus, the maximal value is F(4, 5) = $37.

We can find the composition of an optimal subset by back tracing the computations of this
entry in the table. Since F(4, 5) > F(3, 5), item 4 has to be included in an optimal solution
along with an optimal subset for filling 5 − 2 = 3 remaining units of the knapsack capacity.
The value of the latter is F(3, 3). Since F(3, 3) = F(2, 3), item 3 need not be in an optimal
subset. Since F(2, 3) > F(1, 3), item 2 is a part of an optimal selection, which leaves element
F(1, 3 − 1) to specify its remaining composition. Similarly, since F(1, 2) > F(0, 2), item 1 is
the final part of the optimal solution {item 1, item 2, item 4}.

Analysis
The time efficiency and space efficiency of this algorithm are both in Θ(nW). The time
needed to find the composition of an optimal solution is in O(n).

Memory Functions
The direct top-down approach to finding a solution to such a recurrence leads to an algorithm
that solves common subproblems more than once and hence is very inefficient.
The classic dynamic programming approach, on the other hand, works bottom up: it fills a
table with solutions to all smaller subproblems, but each of them is solved only once. An
unsatisfying aspect of this approach is that solutions to some of these smaller subproblems
are often not necessary for getting a solution to the problem given. Since this drawback is not
present in the top-down approach, it is natural to try to combine the strengths of the top-down
and bottom-up approaches. The goal is to get a method that solves only subproblems that are
necessary and does so only once. Such a method exists; it is based on using memory
functions.
This method solves a given problem in the top-down manner but, in addition, maintains a
table of the kind that would have been used by a bottom-up dynamic programming algorithm.
 16

Initially, all the table’s entries are initialized with a special ―null‖ symbol to indicate that they
have not yet been calculated. Thereafter, whenever a new value needs to be calculated, the
method checks the corresponding entry in the table first: if this entry is not ―null,‖ it is simply
retrieved from the table; otherwise, it is computed by the recursive call whose result is then
recorded in the table.
The following algorithm implements this idea for the knapsack problem. After initializing the
table, the recursive function needs to be called with i = n (the number of items) and j = W
(the knapsack capacity).
Algorithm MFKnapsack(i, j )
//Implements the memory function method for the knapsack problem
//Input: A nonnegative integer i indicating the number of the first items being
considered and a nonnegative integer j indicating the knapsack capacity
//Output: The value of an optimal feasible subset of the first i items
//Note: Uses as global variables input arrays Weights[1..n], Values[1..n],and
table F[0..n, 0..W ] whose entries are initialized with −1’s except for
row 0 and column 0 initialized with 0’s

Example-2 Let us apply the memory function method to the instance considered in Example
1. The table in Figure given below gives the results. Only 11 out of 20nontrivial values (i.e.,
not those in row 0 or in column 0) have been computed. Just one nontrivial entry, V (1, 2), is
retrieved rather than being recomputed. For larger instances, the proportion of such entries
can be significantly larger.

Figure: Example of solving an instance of the knapsack problem by the memory function algorithm

In general, we cannot expect more than a constant-factor gain in using the memory function
method for the knapsack problem, because its time efficiency class is the same as that of the
bottom-up algorithm
 17

6. Bellman-Ford Algorithm (Single source shortest path with –ve weights)

Problem definition: Given a graph and a source vertex s in graph, find shortest paths
from s to all vertices in the given graph. The graph may contain negative weight edges.
Note that we have discussed Dijkstra’s algorithm for single source shortest path problem.
Dijksra’s algorithm is a Greedy algorithm and time complexity is O(VlogV). But Dijkstra
doesn’t work for graphs with negative weight edges.
Bellman-Ford works for such graphs. Bellman-Ford is also simpler than Dijkstra and suites
well for distributed systems. But time complexity of Bellman-Ford is O(VE), which is more
than Dijkstra.
How it works? - Like other Dynamic Programming Problems, the algorithm calculates
shortest paths in bottom-up manner. It first calculates the shortest distances for the shortest
paths which have at-most one edge in the path. Then, it calculates shortest paths with at-most
2 edges, and so on.
Iteration i finds all shortest paths that use i edges. There can be maximum |V| – 1 edges in
any simple path, that is why the outer loop runs |v| – 1 times. The idea is, assuming that there
is no negative weight cycle, if we have calculated shortest paths with at most i edges, then an
iteration over all edges guarantees to give shortest path with at-most (i+1) edges
 18

Bellman-Ford algorithm to compute shortest path

 19

7. Travelling Sales Person problem (T3:5.9),

20
21