TRANSACTIONS ON ENGINEERING, COMPUTING AND TECHNOLOGY V1 DECEMBER 2004 ISSN 1305-5313

An Evolutionary Algorithm to Solve Cryptarithmetic Problem

Abu Sayef Md. Ishaque, Md. Bahlul Haider, Muhammad Al Mahmud Wasid, Shah Mohammed Alaul, Md. Kamrul Hassan, Tanveer Ahsan, Mohammed Shamsul Alam
Abstract Cryptarithmetic problem is an interesting constraint satisfaction problem for which different algorithms have been given. But none of them can solve the problem intelligently using small space and limited time. The paper illustrates the design and implementation of an efficient evolutionary algorithm that can find the solution more quickly taking less memory.

II. STUDY OF THE PROBLEM

Cryptarithmetic is a suitable example of Constraint Satisfaction Problem. Instead of providing description, a Cryptarithmetic problem can be better described by some constraints. The constraints of defining a Cryptarithmetic problem are as follows --Each letter or symbol represents only one and a unique digit throughout the problem. --When the digits replace letters or symbols, the resultant arithmetical operation must be correct. These two constraints lead to some other restrictions in the problem. Consider that, the base of the numbers is 10. Then there must be at most 10 unique symbols or letters in the problem. Otherwise, it would not be possible to assign a unique digit to each unique letter or symbol in the problem. To be semantically meaningful, a number must not begin with a zero. So, the letters at the beginning of each number should not correspond to zero.

KeywordsConstraint Satisfaction Problem (CSP), Cryptarithm,

Cryptarithmetic, Heuristic. Evolutionary Algorithm, Greedy Algorithm,

I. INTRODUCTION

ryptarithm is a genre of mathematical puzzle in which the digits are replaced by letters of the alphabet or other symbols. Cryptarithmetic is the science and art of creating and solving cryptarithms. The worlds best-known Cryptarithmetic puzzle is undoubtedly the puzzle shown in Figure I.

III. FINDING SOLUTION BY SEARCHING

A blind search can eventually find the solutions, if there is any, in a bounded time. Given that the base of the number is 10, there may be 10pn solutions to be checked in the problem space; where n is the number of unique letters or symbols in the problem. As, the problem shown in the Figure 01 has eight different letters S, E, N, D, M, O, R and Y, blind search might have to search at most 10p8=1814400 solutions. If the searching algorithm can check 1000 solutions per second, yet it will take about 30 minutes to get the solution in the worst case. A rule based searching technique can provide the solution in minimum time. Constructing these rules is problem specific and is out of the scope of this paper. The toughness of the problem can be interpreted as less number of valid solutions in relatively huge problem space. The precedence of mathematical operation in the problem indicates a basic estimation of toughness. Again, if there are few letters in the problem, the problem space becomes smaller which makes the problem easy. Given the number of letters in the problem, a problem can be very tough yet. This is because some of the letters or symbols can occur more than once in the problem. These critical letters, the letters that occur more than once in the problem can introduce more restrictions in the problem and reduce the number of valid solutions. An Evolutionary Algorithm will search for solutions in shortest time but the performance will also reflect the toughness of the problem.

S E N D + M O R E --------M O N E Y
Figure I: Cryptarithmetic puzzle This was first introduced by H.E. Dudeney and was first published in the July 1924 issue of Strand Magazine associated with the story of a Kidnappers ransom demand [1]. Modernization, by introducing computers and the Internet, is making quite an impact on Cryptarithmetic and it has already become a standard AI problem because it characterizes a number of important problems in computer science arena.

Manuscript received November 29, 2004. Abu Sayef Md. Ishaque, GrameenPhone Ltd., Telenor affiliated Cellular Phone Company, Bangladesh. (Phone: +880-171-505-583; Fax: +880-2-9882970; e-mail: [email protected]). Md. Bahlul haider, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected]. Muhammad Al Mahmud Wasid, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected]. Shah Mohammed Alaul, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected]. Md. Kamrul Hassan, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected]. Tanveer Ahsan, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected]. Mohammed Shamsul Alam, International Islamic University Chittagong (IIUC), Bangladesh. E-mail: [email protected].

ENFORMATIKA V1 2004 ISSN 1305-5313

494

2004 WORLD ENFORMATIKA SOCIETY

TRANSACTIONS ON ENGINEERING, COMPUTING AND TECHNOLOGY V1 DECEMBER 2004 ISSN 1305-5313

IV. EVOLUTIONARY APPROACH

An Evolutionary Algorithm (EA) is a common term for algorithms that utilize the adaptive behavior modeled after principles of nature [2]. Although definition of Evolutionary Algorithm differs, the more common properties of EAs are that collections of potential solutions to the problem at hand are maintained. These solutions are referred to as the population of a current generation. Each potential solution is called a chromosome. Operations are applied to the current population to produce a new generation that will hopefully contain chromosomes that are better solutions of the problem. This process continues until some threshold value or stopping criterion is met. The new population is produced through the operators on selected chromosomes of current generation. Typically, the chromosomes of the current generation, to whom the operators are applied, are chosen based on their quality. In this way, it is more likely that the offspring chromosomes inherit desirable characteristics of its parents. Some heuristics or fitness functions are used to choose parent chromosomes in a generation. Typically operators are developed from the genetic notation of mutation, recombination (or crossover) and inversion [3]. As an example, a mutation operator can choose two random numbers that may be applied to randomly change some inherited characteristics of the chromosome. Crossover operations involve combining characteristics from two or more parents and placing them in the resulting offspring.

B. Mutation Operation
A mutation operator will randomly generate two different numbers from 0 to 9 and will exchange the positions of the letters in these two indexes. This will correspond to a completely different solution. Y N R M _ O 0 1 2 3 4 5 (a) Y N R M O _ 0 1 2 3 4 5 (b) Y N E M _ O 0 1 2 3 4 5 (c) E D _ S 6 7 8 9 D E _ S 6 7 8 9 D R _ S 6 7 8 9

Figure IV: Three chromosomes after Mutation on Figure 03 Here the only restriction is that the letter that comes at the starting of a number in the problem should not be the starting letter of the chromosome. In the problem of Figure 01 S and M cannot be 0. This is because they are the starting numbers in the problem. A more productive mutation operator ensures that it will not exchange the position of two dont care symbol with each other. Similar cross-over operation can be designed where two chromosomes exchange their attributes. This can be useful in moving faster towards the solution especially at the middle of searching when selected chromosomes adopt some of the good attributes. But for some tough problems it makes the algorithm more susceptible to local minima deadlock. So, cross-over operation may not be included in a general algorithm.

V. FORMULATION OF ALGORITHM A. Encoding solution to chromosome

The initial part of solving a problem in Evolutionary Algorithm is the encoding of solutions into chromosomes. The chromosomes should be simple enough so that operations can be executed easily and descriptive enough to characterize a unique solution. In this problem, letters should be assigned unique digits to characterize a solution. In this paper we will only discuss about the decimal solutions. So, a string of length 10 field by those letters and additional dont care symbol(s) is a good choice. If the base of the number is n then we will use an array of length n. The position of the letters in the string denotes their value. S E N D M O R Y _ _ 0 1 2 3 4 5 6 7 8 9 Figure II: Chromosome In Figure 02, _ (Underscore) is used as a dont care symbol and the numbers just under the letters denote their position in the string as well as their respective values in the solution. Here, 8 and 9 are not used for any letter and the corresponding positions are occupied by dont care symbols. Operators may change the position of the letters and dont care symbol and the chromosome will indicate a completely different solution. Y N R M _ O D E _ S 0 1 2 3 4 5 6 7 8 9 Figure III: Chromosome after operation(s) So, Figure 03 indicates a different solution as the position of letters has been changed.

C. Fitness Function
A Fitness function usually indicates the degree of correctness of a chromosome. An evaluation function can be easily formulated which will calculate the error of mathematical result in the problem. S E N D 9 7 1 6 + M O R E 3 5 2 7 ------------M O N E Y 3 5 1 7 0 ERROR = ABS (35170-(9716+3527)) = 21927 Figure V: Calculating Error in chromosome of Figure 03 The error in the chromosome of Figure 03 is shown in figure 05. Now this can be thought as negative fitness. Chromosome with minimum error is the fittest chromosome in a generation. This algorithm aims to minimize this value. When the value is zero, the solution is found. By only a few modifications in the fitness function we can find the solutions of Cryptarithmetic with more than two operands in the left hand side and with more operators like subtraction, division, multiplication. We can solve the problem SIX + SEVEN + SEVEN = TWENTY. Here we have three operands in the left hand side. So for this problem the fitness function will be as follows ERROR= ABS( TWENTY-( SIX + SEVEN + SEVEN )).

D. Improving Generation through Iteration

Initially we have only one chromosome that is shown in Figure 02. We will first use this chromosome to generate the next generation. By mutation operation this chromosome generates a new generation,

ENFORMATIKA V1 2004 ISSN 1305-5313

495

2004 WORLD ENFORMATIKA SOCIETY

TRANSACTIONS ON ENGINEERING, COMPUTING AND TECHNOLOGY V1 DECEMBER 2004 ISSN 1305-5313

which represents current population. Then we will take some fittest chromosomes from the current generation to produce the next generation. In iterative process, the fittest chromosomes of a generation get the chance to generate the offspring generation. To ensure that the offspring generation is not worse than the current one, the fittest chromosome of the current generation can be added to the next generation. But, this will introduce a deadlock when these fittest chromosomes cannot lead to a solution, even though there exists a solution. So, a random chromosome can be given chance to contribute to the next generation. This modification will avoid the deadlock. The number of chromosomes that will be kept in one generation is problem specific. If we take more chromosomes then we will need more computation and if we take less chromosomes then we might not be able to reach the solution in bounded time. After generating several generations we will find the solution when the error in the chromosome is 0. The solution chromosome of Figure 01 is shown in the Figure 06. O M Y _ _ E N D R S 0 1 2 3 4 5 6 7 8 9 S E N D 9 5 6 7 + M O R E 1 0 8 5 ------------M O N E Y 1 0 6 5 2 ERROR = ABS (10652-(9567+1085)) = 0 Figure VI: The solution chromosome

can say that if we add two n digit numbers and the result is n+1 digit number, then the most significant digit of the result should be 1. Our algorithm can find out that the most significant digit of the result is 1 just after the first iteration. This is because any value other than 1 in the most significant digit in the result will produce larger error values. So our algorithm will reject those chromosomes that have value other than 1 for the most significant digit of the result. According to toughness of particular problem, few parameters like number of chromosomes in each generation, number of chromosomes selected for reproduction and random chromosomes taken to avoid local minima should be set to get solution in a minimal search. Few common unique solution problems are analyzed with 16 chromosomes in each generation and 4 chromosome reproduction within which 1 was taken randomly and statistics shown in Table I have been found in 1000 runs.

Problem
SEND+MORE=MONEY 9567+1085=10652 BASIC+LOGIC=PASCAL 60852+47352=108204 APPLE+LEMON=BANANA 67794+94832=162626 HACK+HACKER=REBOOT 4968+496805=501773 EARTH+URANUS=SATURN 94583+654067=748650 COMET+SATURN=URANUS 61078+298354=359432 BROWN+YELLOW=PURPLE 52813+649981=702794

Min
240 240 336 320 160 1232 288

Max
9248 379520 54288 112560 512000 481584 512000

Avg
1669 10521 5476 2326 49062 96297 87279

Table I: Statistics of searching some standard problems

VI. ALGORITHM
Step 1: Scan the input strings. Step 2: Put the letters or symbols in ARRAY[10] Step 3: If number of distinct letter is less than 10 then fill rest of the indices of ARRAY by dont care symbols. This ARRAY[10] now is our current generation. Step 4: For several times generate two random numbers m, n and swap the contents of index m and n of any one chromosome of current generation and copy this new chromosome to next generation. Step 5: Evaluate the fitness of each chromosome of next generation and choose the best chromosomes. Now these best chromosomes become our current generation. Also include one random chromosome to the current generation. If there is no chromosome with error 0 in the current generation then go to step 4. If one of the chromosomes is found with error 0 then report the solution and exit.

VIII. LIMITATION
This is a greedy algorithm as it is taking the best chromosomes at each iteration and tries to converge to the nearest solution. Like other greedy approaches, it may stick at some local minima. To overcome these local minima we have used a random chromosome at each generation. This will guarantee us that we will be able to overcome the local minima and reach our global minima.

IX. CONCLUSION
This paper concentrated on designing an efficient Evolutionary algorithm to solve cryptarithmetic Problem. Additionally, it illustrates how to plug in techniques of Evolutionary Approach into Constraint Satisfaction Problem. This sort of design can provide efficient solution to a wide range of Constraint Satisfaction Problem or other generic searching problems that could be characterized as a Constraint Satisfaction Problem as well. So, further research should go on to design efficient solutions of different real-life searching problems.

VII. PERFORMANCE
This algorithm is applied on different Cryptarithmetic problems and good result has been observed. For particular SEND + MORE = MONEY problem it checks about 1669 solutions on an average whereas blind search requires about 10p8/2=907200 solutions to be checked on an average. The number of solutions to be checked varies because of the random property of this algorithm. If we take a closer look at the algorithm we will find out why this algorithm can find the solution quickly. In the problem of Figure 01 we see that we are adding two four digit numbers (SEND+MORE) and the result is a five digit number (MONEY). From our mathematical experience we

REFERENCES
[1] http: // www.geocities.com / Athens / Agora / 2160 / index.html [2] Evolutionary Algorithm Approach for Symbolic FSM Traversals, IEEE Pacific Rim Conference on Communications, Computers and Signal Processing (PACRIM), August 26-28, 2001, pp. 506-509, (with R. Drechsler). [3] R. Drechsler, Evolutionary Algorithm in VLSI CAD, Kluwer Academic Publishers, Boston Hardbound, ISBN 0-7923-8168-8, May 1998.

ENFORMATIKA V1 2004 ISSN 1305-5313

496

2004 WORLD ENFORMATIKA SOCIETY