See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/262281816

Dynamic programming from an excel perspective

Article April 2013

CITATIONS READS

0 558

3 authors:

Ranette Halverson Richard Simpson

Midwestern State University Midwestern State University
20 PUBLICATIONS 30 CITATIONS 19 PUBLICATIONS 18 CITATIONS

SEE PROFILE SEE PROFILE

Catherine Stringfellow
Midwestern State University
30 PUBLICATIONS 156 CITATIONS

SEE PROFILE

All content following this page was uploaded by Richard Simpson on 13 August 2015.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 DYNAMIC PROGRAMMING FROM AN EXCEL PERSPECTIVE

Ranette Halverson, Richard Simpson, Catherine Stringfellow

Department of Computer Science
Midwestern State University
ranette.halverson, richard.simpson, catherine.stringfellow@mwsu.edu

ABSTRACT
Many students have difficulty understanding the concept of dynamic programming, a problem
solving approach appropriate to use when a problem can be broken down into overlapping sub-
problems. In addition, computer science majors seldom take a course in which a spreadsheet
application is taught or used, although spreadsheet applications can be a great tool in visualizing
and solving problems. This paper attempts to address these issues by providing examples of
problems that can be solved in a spreadsheet environment using dynamic programming in
various computer science courses.

INTRODUCTION
There is a well-known approach to solving certain algorithms called Dynamic Programming (DP),
an unusual term that resulted from the fact that the 1950 Secretary of Defense hated
mathematics [2]. It is primarily a bottom-up method that has been around since the 1940s
used for developing a recursive, top-down solution for many problems. It is appropriate for use
with problems that can be solved by breaking the problem down to overlapping sub-problems.
The popularity of this approach is heavily due to its efficiency. A solution having an exponential
complexity for a top-down recursive algorithm may have a polynomial complexity with the
associated dynamic programming method. When students are first introduced to this approach
as a problem solving method, they are often uncomfortable or at least have trouble visualizing
the process. The teaching method discussed in this paper attempts to address these educational
road-bumps. A collection of problems from very simple to somewhat complicated are solved in
a spreadsheet environment using the dynamic programming approach to demonstrate its
algorithm visualization value. A positive side effect is that computer science majors become
better acquainted with spreadsheet programming.

BACKGROUND
There has been previous work in combining MS Excel and Dynamic Programming (DP)
particularly in the area of operations research modeling. Jensens website Operations Research
Models and Methods is a useful site from which to download Excel add-ins that solves a variety
of problems using dynamic programming. This software is in support of his book of the same
name [4]. Professor Mark Gerstein in his Yale bioinformatics lab has been using similar
techniques for visualizations of DNA sequence matching (basically a variation of the longest
common subsequence problem) as well as other problems [3]. Consequently many of the
techniques that will be discussed may well have been used in a variety of educational settings
although it doesnt appear to be widespread in dynamic programming. It is the purpose of this
paper to demonstrate educational applications of Excel to DP. It is not the purpose of this paper
to explain the dynamic programming algorithmic approach.
SIMPLE EXAMPLES
Algorithm designers generally consider problems suitable for dynamic programming if the
problem can be modeled by top-down recursive representations of sub-problems, which have
an optimal substructure. The efficiency of the approach is greatly increased if the sub-problems
have a large amount of overlap. This will become clear within the context of some of the
examples.

We begin with some very simple examples that many may not even consider to be DP due to
their simplicity. Skipping probably the simplest, which is the calculation of factorial(n), we will
start with the calculation of the nth Fibonacci number. Many computer science students have
written this function recursively as one of the traditional problems presented when introducing
the concept of recursion. Due to the exponential complexity one would not usually use the
algorithm shown in Figure 1 so we turn out attention to the algorithm in Figure 2. This
algorithm is a bottom-up solution to the top-down recursion solution shown in Figure 1. In
other words, it is DP in a very simple form. The resulting complexity is O(n) not O(cn), a
significant improvement.

int Fib( int n)

int Fib( int n) {
{ A=new int[n+1];
if (n<3) return 1 A[1]=A[2]=1;
else return ( Fib(n-1)+Fib(n-2) ); for(int i=3 ; i<=n ; i++)
} A[i] = A[i-1] + A[i-2];
return A[n];
}
Figure 1: Recursive Fibonacci.
Figure 2: Iterative Fibonacci.

Switching from C++ to a spreadsheet environment (MS Excel, in this case) one can quickly build a
simple visualization of the Fibonacci process. Place the value 1 in cells A1 and B1 and the
formula A1+B1 in cell C1. Copying C1 to the cells to the right immediately constructs the
Fibonacci sequence in a bottom-up DP manner. This is, of course, a very simple example in
which the value in each cell is determined by the sum of the previous two cells.

Moving from this one-dimensional setting to two-dimensional provides a more interesting

scenario. Suppose we want to determine a particular coefficient of the binary expansion, i.e.
(a+b)n. The solution is a cell in Pascals Triangle. The nth row, kth column of this triangle is given
by the recursive definition

Pascal(n,k) = Pascal(n-1,k) + Pascal(n-1,k-1)

Pascal(n,0) = Pascal(n,n) = 1

Since writing a program that returns the value of a

cell in the triangle recursively is computationally
exponential, other approaches are often used. As
with the Fibonacci array we can evaluate a cell by
filling required portions of the array bottom-up. This
approach, which is the usual procedure an instructor
applies when the triangle is written on the board, is Figure 3. Pascals Triangle.
another dynamic programing example. If we can develop the triangle in a spreadsheet, one can
see the triangle and its values easily. Enter the value 1 in the cell A1 and also in the row 1 and
column A of the triangle. (See Figure 3.) Next, place the formula =A2+B1 in cell B2 and copy to
the remaining cells; this builds the triangle. Although its too fast to see as it develops, the cells
are filled from the upper left to lower right. The evaluation looks to be dynamic programming
even if the software within Excel works differently. The main point is that it takes very little
Excel programming, in this case at least, to build and evaluate a dynamic programming table.

MORE COMPLEX EXAMPLES

Due to the simplicity of the previous problems, the dynamic programming process might not be
clear to those unfamiliar with the process. To remedy this, the well-known Shortest Common
Subsequence problem will be addressed [1]. Although more complex variations of this problem
are heavily used in bioinformatics and DNA string matching, we use the simplest version for this
paper.
A subsequence of a sequence X (think string, if you must) is a sequence obtained from X by
removing 0 or more of the characters of X with the remaining characters kept in the same order.
For example the sequence abrdfredffd is a subsequence of X = abfrdfdvredffgfd. The
subsequence is obtained by removing the bold letters of X. Define a Common Subsequence of
sequences X and Y as any sequence that is a subsequence of both X and Y. There may be many
subsequences of X and Y. A subsequence of this set that has maximal length is called the
Longest Common Subsequence (LCS), and there may be more than one of these. The Common
Subsequence Problem is one in which we seek to determine a least one of the maximal
sequences for two given sequence X and Y. To solve this problem requires some definitions, an
approach that common in dynamic programming. Let Xi represent the first i characters of X and
Yi the first i characters of Y. Assuming that Xm = x1x2x3xm and Yn = y1y2y3yn and let Z=
z1z2z3zk be any LCS of X and Y. Now there are three situations that require consideration. Note
LCS is defined recursively in terms of smaller sub-problems.

1. If xm=yn then zk=xm=yn and Zk-1 is the LCS of Xm-1 and Yn-1
2. If xmyn then zk xm implies that Z is an LCS of Xm-1 and Y
3. If xmyn then zk ym implies that Z is an LCS of X and Yn-1

Combining these in a recursive definition gives:

where c[i,j] is defined to be the length of the LCS of prefixes Xi and Yj . Note, the length of the
LCS is defined, not the actual sequence.

The above definition specifies the value of the cells in a 2-dimensional table. For example,
placing two strings, e.g. BDCABA and ABCBDAB in the indicated row and column respectively,
the initial table is built as shown in Figure 4. This table is filled with the two strings and the
initial 0 values in row 3 and column C. The cells in the D4:I10 rectanuglar area represent the
values for c[i,j]. In particular c[1,1] is stored in cell D4 and c[2,3] is stored in F5 and so forth. To
evaluate this table enter an Excel function for cell D4 and copy to the remaining cells. This
function is the previously defined formula for c[i,j]; that is, IF(D$2=$B4,C3+1,MAX(D3,C4)). (It is
important to note that many computer science students have little experience with Excel. They
occasionaly build tables and charts for papers and such, but seldom take a course that trains one
in MS Office tools. Thus, a simple review of absolute addressing and functions such as IF is
probably called for.) Placing this formula in cell D4 and copying to the remaining cells generates
the table on the right of Figure 4. From the lower right hand cell in the table one can see that
the length of the longest common subsequence is 4. The length of the LCS is increased by 1
every time two of the characters from the strings match. Note that cell F6 is obtained by
incrementing E5 by 1, a result of the fact that both F2 and B6 contain a C. The sequence of
characters that generate this behavior is the LSC, i.e. BCBA in this case.

Figure 4. Longest Common Subsequence Tables.

As a final example we show that it is possible to push this concept too far. There are problems
for which a DP solution is complex enough that retrofitting the algorithm within Excel becomes
overly complicated. An example of this is the well-known Optimal Binary Search Tree problem
[1]. The problem is to construct a binary search tree in such a way as to optimize search time,
given the probability of searching for each key. In other words, the tree should produce the
minimal expected search cost. In the interest of brevity, let us consider the DP recurrence
relation that results from this problem and omit the reasoning process that generated the
relation. Solving this problem requires the constructions four tables (tables) e, weight (w), root
and probability (Figure 5). The cells c(i,j) in the e array contains the cost of searching the sub-
trees containing nodes i to j. The weight array contains the sums of the probabilities of the
indicated nodes. The final answer is contained in the root array which indicates the structure of
the optimal tree. The recurrence for e is

{ { }

where
From the recurrence relation, the tables shown in Figure 5 are built. The w array and the p,q
array , which contains the initial set of probabilities of accessing the nodes, are easily
established leaving the e array and the root array to be calculated. DP is applied to the e array
and during the process the root array is filled.

To clarify the problem, first note that the e array is constructed from the lower right to upper
left with the last step being the determination of cell(5,1) which is 2.75. Each cell is evaluated
by used the previously calculated values below and to the right of the cell as well as using the
corresponding weight cell from the w array. For example cell(6,2) which is 1.2 can be
determined as follows. (Figure 6.)

Cell(6,2) = min { cell(7,2) +cell(6,5), cell(8,2)+cell(6,4),cell(9,2)+cell(6,3)} + .5 (from the w array)

Cell(6,2) = min { .7+.05, .4+.3, .1+.6} +.5 = 1.2

Figure 5. Optimal Binary Tree Tables.

This process is straightforward to write in C++, but can we

do it in Excel? It is not as easy as one might think.
Developing an Excel formula for the cells that calculate the
formula required in cell(6,2) is not obvious. Cell(5,1)
requires a minimum applied to four sums, the next
diagonal of numbers require 3 sums and so on. One could
write each of these and copy diagonally but this is time
consuming and the resulting DP visualization becomes
confusing and less satisfying to a student. One reasonable
solution is to write an Excel user defined function (UDF)
using Visual Basic (VBA). VBA is the built-in macro
language used by Excel. To investigate the effort required
to write a UDF, we wrote the function in Figure 7. This
Figure 6. Table Processing
function can then be copied to each cell and the values
will be calculated as desired. However this method causes
additional problems. Most students and perhaps the faculty as well, have little advanced
knowledge of Excel. (That is true in our department.) Spending time learning the system and
methods for building VBA functions in Excel might be considered getting off-track of the goals of

Figure 6. Table Processing.

 this project. There is also another problem within Excel itself. Within the function definitions,
code was included to modify the root array as it ran. The values in the root array result from the
minimum sums in the e array. This is not
possible as Excel does not allow cell Function minbin()
Dim Row As Integer
modification side effects within functions. Dim Col As Integer
Alternatively, one could write a separate Dim Wid As Integer
function to be used in the root array table Dim Count As Integer
Dim Min As Double
but this would just be a modified copy of Dim Val As Double
minbin() . Whats the solution? An obvious Min = 100000
Row = Application.Caller.Row ' Get row of calling cell
method that would work efficiently is to Col = Application.Caller.Column 'Get column of calling cell
write the entire algorithm as a procedural Wid = Cells(Row, 7) - Col + 1
For Count = 1 To Wid
macro in VBA and have it fill in all the cells. Val = Cells(Row + Count, Col) + Cells(Row, Col + Wid + 1 -
Of course, in this case why not just use C++ Count)
in the first place? The moral of the story is If Val < Min Then
Min = Val
that some DP problems are just too End If
complicated to easily implement in Excel, Next Count
minbin = Min + Cells(Row, Col + 7)
especially if the original reason for doing so Application.Volatile ' Force recalculation of cells.
was a better understanding and End Function
visualization of dynamic programming
Figure 7. Excel function minbin()
tables.

CONCLUSION
Over the years it has become clear to the authors that tools such as Excel (or other spreadsheet
software) are not appropriately appreciated by many computer science students. Most never
take a course in Excel and hence are quite unaware of its advanced features beyond simple
applications.

We make a conscious attempt to embed Excel into many of our courses so that the students will
at least be exposed to spreadsheets concepts. In addition, it is a good tool to visualize table
construction algorithms. In particular, it has been used in Data Structures to graph data, in
Introduction to Computing for Science Majors to build tables such as Fibonacci and Pascal and in
Bioinformatics in which the Longest Common Subsequence Algorithm is always investigated. In
other words, since most computer science programs do not include a course in spreadsheets, it
is desirable to include it throughout a computer science program. Dynamic programming
techniques with their associated visualization are one of the most interesting, as well as
educational applications, of a spreadsheet in a computer science program.

REFERENCES
[1] Cormen T, Leiserson C, Rivest R, Stein C. Introduction to Algorithms. 3rd ed. Cambridge, Mass: MIT
Press; 2009.
[2] Dreyfus S. Richard Bellman on the Birth of Dynamic Programming. Operating Research.
2002;50(1):48-51.
[3] Gerstein M. Gerstein Lab. [Internet]. 2011 [cited 2012 November 1]. Available from:
http://www.gersteinlab.org/.
[4] Jensen PA, Bard JF. Operations Research Models and Methods. Hoboken, NJ: John Wiley & Sons, Inc;
2003.

View publication stats