next_inactive up previous
Up: Course Projects

CPSC 601.73 Project
Evolution with Introns and Exons

Jie Gao
Department of Computer Science
University of Calgary
gaoj@cpsc.ucalgary.ca

Abstract:

This paper is the report of my project for the course Biological Computation. Based on the fact that there are large quantities of genes whose functions are unknown, and enlightened by the paper [WBM03], I proposed a new Genetic Programming approach with introns and exons that allows part of the genes being disabled. I implement two different program expression approaches, namely byte sequence and tree structure. The tree structure expression performs well in evolution. Some experiments are done to compare the traditional Genetic Programming with this new approach. The result shows my new approach can really increase the adaptability of Genetic Programming.


Introduction

In computer science, evolutionary computation is an important computational paradigm that comes from the nature. It tries to use evolution mechanisms existing in the nature to solve computer problems. Therefore new discoveries and hypothesis on the natural evolution could give new ideas to evolutionary computation.

According to the known facts, in creatures' DNAs there are large quantities of genes that are not ``translated'' by the creatures into the phenotypes, and they are sometimes called ``garbage genes''. But most people, including me, think they are not garbages at all. Their functions are simply unknown currently. So many people use the term introns to refer to such function-unknown genes instead of ``garbages''. Correspondingly, the known genes are called exons. Many hypothesis are made on these introns. People think they could be meta data of the exons, or the records of the whole evolution histroy[Kam].

In January, a paper in Nature with the title Loss and Recovery of Wings in Stick Insects[WBM03] reported a new discovery in evolution -- some species can lose a complex organ and get it back some time later. This phenomenon may happen for several times during the evolution. It amazes scientists a lot because it reveals a fact that some complex functions of a species can be temporarily disabled and latent in the genes and be triggered by the environment again some generations later to come back to the species.

This discovery interests me greatly. Associating the introns facts and the phenomenon revealed by the paper, I come to the hypothesis that these genes are turned from exons into introns -- thus disabled temporarily and still in the genes, and later turned from introns back to exons -- enabled again. It seems that such a mechanism can probably increase the adaptability of creatures, especially in an ever-changing environment.

A natural thought is to test whether such a mechanism can result in increase of adaptability in evolutionary computation. Therefore my idea for this project comes into being -- a Genetic Programming with introns and exons.

In this report I explain the experiments on the Genetic Programming with introns and exons, and show the results I got: Section 2 presents the problem selected for the Genetic Programming to solve; the implementation is discussed in Section 3; Section 4 explains the experiments I made, as well as the results I got from the experiments; finally in Section 5 and 6, the report is concluded and future work is indicated.


Problem to Solve -- Robot in a Maze

Scenario

The problem for my Genetic Programming with introns and exons to solve is ``Robot in a Maze'': A maze is an array of grids (see Figure 2 on Page [*]). The grid types can be path (the robot can go through) or wall (the robot can not go through). The maze has a start point and a finish point (marked by S and F in the maze). The robot's task is to find a path from the start point to the finish point. Particular programs are used to express different moving plans of the robot. These programs are evolved by Genetic Programming.

Capabilities of the Robot

The robot has various capabilities, including sensing the environment, moving and doing simple logic and flow control operations. These elements of capabilities consist the operations of the programs that will be evolved in Genetic Programming.

The detailed capabilities of the robot, or the available operations in the program, are listed in Table 1. Each of them has a boolean return value that can be an operand of the logic operations. Nop is a special operation that does nothing but generate a random boolean value as the return value. If some of one operation's operands are missing, Nop will be appended to make the operation valid.


Table 1: Operations of the robot in a program
Name Operand No. Return Value Description
E 0 Whether the grid to the east is a path grid
N 0 Whether the grid to the north is a path grid
S 0 Whether the grid to the south is a path grid
W 0 Whether the grid to the west is a path grid
NE 0 Whether the grid to the northeast is a path grid
NW 0 Whether the grid to the northwest is a path grid
SE 0 Whether the grid to the southeast is a path grid
SW 0 Whether the grid to the southwest is a path grid
East 0 Succeed? Move to the grid to the east
North 0 Succeed? Move to the grid to the north
South 0 Succeed? Move to the grid to the south
West 0 Succeed? Move to the grid to the west
NorthEast 0 Succeed? Move to the grid to the northeast
NorthWest 0 Succeed? Move to the grid to the northwest
SouthEast 0 Succeed? Move to the grid to the southeast
SouthWest 0 Succeed? Move to the grid to the southwest
Not 1 Logic not of the operand
And 2 Logic and of the operands
Or 2 Logic or of the operands
If 3 The first operand If the first operand is true, execute the second operand. Elsewise execute the third operand
Nop 0 Random No operation is executed

The Program

A program of the robot's moving plan is consisted by the operations listed in Table 1. Because the operations have only one type of return values, and they all have return values, all operations are compatible. Thus all kinds of operations can be an operand of another operation. A program is organized in a hierarchical structure.

The introns and exons are labelled by a particular flag -- an enable flag on each of the operations. One operation is an exon if it is enabled. Otherwise it is an intron. In the hierarchical structure, if one operation is an intron, all its operands and operands' operands are introns.


Implementation of Genetic Programming with Introns and Exons

I have tried two different approaches to express the programs. One is encoding the programs by byte sequences, the other is using tree structures to express programs. I implemented Genetic Programming with introns and exons in both approaches and they got different performances in evolution. The following subsections discuss the two approaches and the difference between them.

Byte Sequence Encoding Approach

Expression

In this approach each operation is encoded by a binary integer. Since the total number of different operations is not large, one byte is enough for this binary integer, together with a bit for enable flag. In my actual implementation, the operations are encoded in the way described in Table 2. In this way, not only the operation code itself, but also the related information of the operation is encoded in the byte. The length of the byte sequence is not fixed so that the programs can be in various complexity. But in actual implementation, there is a hard limit -- 1024 bytes so that the size of the byte sequence is practical in implementation.


Table 2: Byte encoding of the operations
7 6 5 4 3 2 1 0

 

Bit Usage
0 - 2 Operation code
3 Operation type:
  0 -- Detection of grids;
  1 -- Movement.
  Ignored if bits 4-5 are not all zeros.
4 - 5 Number of operands encoded in binary integer
6 Nop flag
7 Enable flag

The program is first expressed in operator-postfix expression, then converted into a byte sequence. The postfix expression ensures that the hierarchical structure of the program is preserved when it is converted into a flat byte sequence. And the postfix expression also makes the program easier to interpret, as is discussed below.

Evolutionary Operations

The evolutionary operations on the programs include mutation and crossover. The mutation operation will randomly select one byte in the byte sequence, and then randomly choose one of the three operations: deleting this byte, inserting a random byte in front of this byte, or changing this byte with another random byte. Here the random byte is verified to express an existing operation in Table 1. Crossover takes place on two byte sequences (programs) $A$ and $B$: Randomly select one byte in $A$ and another byte in $B$, then concatenate the subsequence form the beginning of $A$ to the selected byte with the subsequence from the selected byte in $B$ to the end of $B$.

The evolutionary operations do not check the structure of the new program, because the execution mechanism ensures all combinations of operations in Table 1 are executable.

Execution

A byte sequence is first converted into an execution table in which low-level commands are recorded for execution -- ``compiling'' the program before execution. In the execution table each low-level has an address, a operand address list and a value (see Figure 1). All operations in Table 1 except If have corresponding low-level commands. If operation is converted to combination of several low-level commands including TSTJ (test a value in one address and jump to another address according to the value tested) and JMP (jump to another address). The postfix expression makes the conversion easier. A stack is enough for conversion. While converting one operation, if no enough operands exist, one or more NOP low-level command are inserted to make the operation has enough operands.

Figure 1: Low-level command format
\begin{figure}\begin{center}\scriptsize\tt \begin{tabular}{\vert c\vert c\vert c... ...erand 1&Operand 2&Operand 3&Value\\ \hline \end{tabular}\end{center}\end{figure}

After the conversion, the execution table is executed over a maze to simulate the movement of the robot. The execution is in iterations until the robot reaches the finish grid, or the execution step limit is reached.

Tree Structure Approach

Expression

The tree structure is a ``standard'' expression in Genetic Programming. In this approach, each operation is expressed by a node in a tree, and its operands are children nodes. The children of a node are ordered so that they can express the operands of an operation. A tree directly represents the hierarchical structure of a program.

In each node, the operation and its return value are stored, together with an enable flag. The enable flag decides the gene type (introns or exons) of the whole subtree starting from this node.

In actual implementation, the tree has a node index to list all nodes in this tree so that the evolutionary operations can randomly choose a node easily. Like the byte sequence approach, the number of nodes in a tree is not fixed but limited to 1024 for practical reasons in implementation.

Evolutionary Operations

The evolutionary operations in this approach are also mutation and crossover. In mutation a node in the tree is randomly selected (from the node index), one of the two operation is randomly executed: Flipping the enable flag of this node, or replacing the subtree starting from this node with a randomly generated and validated tree (could be an empty tree). In crossover two parent trees $A$ and $B$ contribute to the new offspring. A randomly selected subtree in $A$ is replaced with another subtree randomly selected from $B$.

Execution

The execution of a tree is in an ``interpret'' way other than the ``compile and execute'' way in the byte sequence approach. The execution starts from the root node of the tree and recursively follow the whole tree structure. Like the byte sequence approach, this program tree is executed in iterations until the solution of the maze is found, or the execution steps exceed the execution limit.

Comparison of the Two Approach

The two approach both provide the expressions of the programs, evolutionary operations that achieve Genetic Programming, and simulations of the robot moving in the maze. But they have different characteristics and got different results in my experiments.

Characteristics

The byte sequence approach simplifies the evolutionary operations. Mutation and crossover on the byte sequence do not need to consider more about the inside of the sequence. Comparing to the tree approach, the usage of pointers is avoided. Thus in implementation this approach is faster and less error-prone. The mechanism in evolutionary operations and execution get faster speed than the tree structure.

The tree structure approach is more clear in the hierarchical structure. This approach is the ``standard'' way in Genetic Programming and it is tested by various Genetic Programming implementations. But the evolutionary operations and execution are more complex than the byte sequence approach.

Different Results

Both approaches are implemented and tested. But they perform differently in the evolutionary behaviours. The byte sequence approach get bad results in evolution. The fitnesses over generations cannot converge. It's more like random behaviours other than evolution. The tree sequence works well in evolution. The species can increase their fitnesses over generations towards the final goal.

The failure of the byte sequence approach is analyzed. The main reason is that the evolutionary operations in this approach cannot keep a relatively stable structure of the program. Good parts in one species can not be transferred to its offsprings via evolutionary operations. Although the execution of the byte sequence ensures every byte sequence is executable, it causes the simplified evolutionary operations changing the hierarchical structure greatly, especially when operations like If is mutated. Thus the evolutionary operations in the byte sequence approach are not really evolutionary.

 

Further experiments are done based on the tree structure approach. The experiments and results are described in the next section.


Experiments and Results

The following experiments and results are all based on the tree structure implementation of Genetic Programming with introns and exons. To test whether this Genetic Programming approach with introns and exons can get better results than the traditional Genetic Programming, a ``standard'' Genetic Programming is also implemented as the reference. In this traditional GP implementation, the programs are in the same tree structure as the one described above, but the enable flags are ignored so that it is a ``standard'' way.

In the subsections below, the term new approach refers to my Genetic Programming with introns and exons, and old approach refers to the traditional Genetic Programming approach.

The Experiments

I hypothesize that the introns and exons mechanism can increase the adaptability of the species in evolution. Therefore the experiments should be able to test and compare the adaptability of the new and old approaches. For this purpose, the programs are evolved continuously on three different mazes. At the beginning of the experiment, generate the first generation randomly and evolve them on the first maze. If one of the programs can lead the robot to move from the start grid to the finish grid -- solution found, change current maze with a second maze, and evolve the existing programs directly in this new maze. After the solution of the second maze is found, replace this maze with a third maze and continue the evolution directly in that maze. When the solution of the third maze is found, stop the evolution and the experiment ends.

Based on this design of the experiment, I have chosen the parameter values as follows (they apply to both the new and old approaches):

Figure 2: Mazes used in experiments
\includegraphics[scale=0.7]{map6x6.1.eps} \includegraphics[scale=0.7]{map8x8.1.eps}
(a) 6$\times$6 maze (b) 8$\times$8 maze
\includegraphics[scale=0.7]{map10x10.1.eps}
(c) 10$\times$10 maze
Black: Wall grid   S: Start grid
White: Path grid   F: Finish grid

In the experiments, I define one test as:

To get convincing results, I ran my experiment for 100 tests. The next subsection shows the results.

Results

After running 100 tests, the direct results are the generation numbers spent in each maze to find the solution. The generation numbers of the first solution programs are shown in Table 3. If a generation number in it is 100, it indicates the evolution fails finding a solution within limited generations in corresponding maze. From this table, as well as the statistics in Table 4, we can find the new and old approaches' performances are almost the same in maze 1 and maze 2. But in maze 3 there is great difference between the two approach: The majority of the new approach tests can find a solution in maze 3 while almost none of the old approach tests (except one) can find a solution in maze 3.

Table 3: Result table$^1$ from 100 tests
New Approach Old Approach
6$\times$6 8$\times$8 10$\times$10 6$\times$6 8$\times$8 10$\times$10
4 47 49 0 41 100
4 24 100 0 75 100
0 27 86 2 35 100
0 24 57 0 18 100
0 19 49 0 28 100
0 20 100 0 24 100
0 100 67 2 49 100
0 60 72 0 64 100
0 33 100 0 28 100
0 22 100 0 26 100
0 23 100 0 91 100
0 12 52 0 42 100
0 100 39 0 50 100
0 42 53 0 18 100
0 39 69 0 48 100
0 100 42 0 14 100
0 12 24 2 33 100
1 100 83 0 38 100
0 30 100 0 31 100
2 38 65 0 39 100
0 44 100 0 27 100
0 21 76 0 100 100
0 11 37 0 28 100
0 31 57 0 35 100
1 100 81 0 47 100
0 22 80 0 29 100
0 39 100 0 42 75
0 14 68 0 26 100
3 28 43 2 48 100
0 24 100 0 57 100
0 15 78 0 25 100
0 17 100 0 39 100
0 32 100 0 30 100
0 66 47 0 28 100
3 55 100 0 40 100
0 27 56 0 24 100
0 23 81 0 54 100
1 100 62 0 19 100
0 31 100 0 25 100
0 100 54 0 26 100
0 7 40 0 35 100
2 17 100 0 43 100
0 100 100 0 32 100
4 23 80 0 26 100
2 58 65 0 26 100
0 100 100 0 34 100
0 26 100 0 24 100
0 49 100 2 13 100
0 33 64 0 49 100
0 22 100 0 45 100
New Approach Old Approach
6$\times$6 8$\times$8 10$\times$10 6$\times$6 8$\times$8 10$\times$10
0 24 45 0 20 100
3 48 76 0 34 100
0 55 89 0 34 100
1 37 55 0 50 100
0 25 69 0 26 100
0 14 47 0 38 100
0 30 83 0 28 100
0 20 73 0 22 100
1 49 46 0 38 100
0 18 34 0 36 100
0 23 51 0 23 100
1 100 71 0 34 100
0 30 100 0 44 100
0 100 60 0 43 100
0 56 47 0 25 100
3 21 69 0 19 100
0 26 75 0 34 100
0 24 78 0 27 100
1 37 100 0 34 100
1 31 100 0 100 100
1 31 74 0 43 100
0 16 60 1 36 100
0 11 100 0 100 100
2 100 72 0 34 100
0 100 86 0 100 100
0 53 41 1 69 100
6 16 100 0 27 100
0 100 100 0 57 100
0 35 68 0 26 100
0 67 100 0 35 100
0 100 91 0 100 100
0 18 44 0 32 100
0 21 52 0 39 100
0 30 61 4 100 100
0 32 50 0 28 100
0 27 69 1 23 100
1 24 100 0 24 100
3 14 63 0 20 100
1 29 45 0 25 100
3 15 65 0 100 100
0 33 65 0 20 100
0 28 64 0 51 100
0 19 66 0 37 100
0 24 56 0 100 100
0 25 100 0 88 100
0 21 38 0 18 100
0 13 46 1 22 100
0 19 71 0 100 100
0 35 50 0 100 100
0 36 77 1 20 100


Table 4: Statistics from result table
  New approach Old approach
Total tests 100 100
Success in maze 1 100 100
Success in maze 2 85 90
Success in maze 3 72 1

The averages of the highest fitness in each generation in the 100 tests are shown in Figure 3, 4 and 5 respectively for maze 1, 2 and 3. They can convince us again what Table 3 and 4 show. In Figure 3 and 4, the two curves are very close to each other, which shows the performances of the two approaches are almost the same. But in Figure 5, the curve for the new approach is higher than the one for the old approach. This again clearly shows the new approach get better performance.

Figure 3: Average fitness of fittest species in maze 1
\includegraphics[scale=0.8]{compare1.eps}

Figure 4: Average fitness of fittest species in maze 2
\includegraphics[scale=0.8]{compare2.eps}

Figure 5: Average fitness of fittest species in maze 3
\includegraphics[scale=0.8]{compare3.eps}

In the experiment, the programs are evolved in three different mazes, which stand for various environments for the evolution. From the result of the experiment on the three mazes, we can tell that if the population get fit in one maze, they can get fit in another maze faster in the new approach than the old one. Thus the adaptability of the new approach is better than the old one.

Analysis of Introns

The Genetic Programming with introns and exons gets better results than the ``standard'' Genetic Programming. The only difference between the two approach is the mechanism of introns and exons, which allows a partition of genes being disabled temporarily and enabled again some generations later. So analysis on how the introns affect the evolution is necessary.

Due to the large number of tests, I only choose the new approach in one test to analyze. The one I have chosen is No. 60 test, whose generation numbers in the three mazes are 0, 18 and 34 respectively. I recorded all random number seeds in the tests. So I can replay a particular test for many times with exactly the same results.

Before tracing down generation by generation, I gathered some statistics from this test. They show totally how many programs are in the elite groups, and how many times an enable flag is flipped. The numbers are listed in Table 5.


Table 5: Statistics of new approach in No. 60 test
(a) Test information: Generation numbers
Maze 6$\times$6 8$\times$8 10$\times$10
Gen. No. for finding the solution 0 18 34
Actual Gen. No. evolved 4 22 34

 

(b) Statistics about evolution
Program type Number Percentage
Produced by evolution2 251937 100%
With enable flag flipping 18297 7.263%

 

(c) Statistics about elite groups
Program type Number Percentage
In elite groups from all generations 63063 100%
Contain introns 32021 50.776%
Average percentage of introns in a program 6.579 %

From the statistics in Table 5(b) we can find actually flipping enable flags takes place at a low frequency. But in the elite group about half of the programs contain introns. This shows that the introns will not decrease the fitness of the programs. And in the programs, the introns do not hold a very high percentage comparing to the exons. On average the intron percentage is only 6.579%. Comparing to the reality, it is a rather small percentage. The amazing thing is that such a small amount of introns are able to result in greatly difference in the adaptability.

Recalling my hypothesis in Section 1, the expected phenomenon that increases the adaptability of the evolution is that some exons are mutated into introns in one generation, and in later generations part or all of these introns are mutated into exons again and they contribute much to the increase of the fitness. Therefore more detailed analysis should be tracing some program subtree's origins in some generations before to see if such phenomenon happens.

To analyze such phenomenon, a family graph of the programs is needed. However, the total generation number on the three mazes is 63 and the family graph should have about $(2^{64}-1)$ edges, which means the family graph would be enormous huge if I were able to generate it. I am able to analyze only part of the family graph. I choose the fittest program in the final generation, and generate the ancestor tree for 20 generations (see Figure 6). The ancestor tree for 30 generations is over 2.1 G Bytes (the size when I kill my ancestor tree generator), and I am not able to analyze it at all.

Figure 6: Part of the ancestor tree
\includegraphics[scale=0.85]{tree.eps}
About 1/650 of the ancestor tree generated from the fittest program in the final generation. The numbers are the program IDs used in the experiment. To simplify the tree, if a node has two identical branches one of them is omitted.

This tree is still very large. I am only able to follow the evolution backward through the ``major contributors'', which contribute most of the program nodes in its direct children. E.g. if program $A$ and $B$ crossover to produce program $C$, and one of $A$'s subtree is replaced by only a single node from $B$, I will concentrate on the transition from $A$ to $C$. By this way, it is possible to trace down from the root node in the tree to a bottom node. However, through the analysis here, I only see sometimes a single node is changed from exon to intron. No further complex introns and exons transitions are observed.

After analyzing this ancestor tree, I think the reason why I fail to observe the expected phenomena is that the transitions from introns to exons most take place when the programs are moved from one maze to another. The last 34 generations are all in maze 3 but I only have a 20-generation tree. So I can not see what happens immediately after changing the mazes.

To see a whole evolution process from the very beginning to the final generation, I generate the ``Father Line'' and ``Mother Line'' of the fittest program in the final generation. The Father Line contains only the father of the program, and father's father, and so on.... Similarly, the Mother Line contains only mothers and mothers' mothers. Table 6 shows the two ``lines''.


Table 6: ``Father Line'' and ''Mother Line''
Father Line
Program ID Intron size
331009 1
321604 1
312546 1
307127 1
300763 0
294446 0
288409 0
283226 0
276550 0
269501 1
258747 1
252770 0
244804 0
239268 0
232546 0
219722 0
211662 0
202904 0
197453 0
194128 0
189839 0
183012 0
174951 0
164590 0
157700 0
149204 0
142889 0
133150 0
117350 0
105446 0
101136 0
94579 0
90302 1
80823 1
72793 1
66013 1
62446 0
58326 0
52441 0
46953 0
42819 0
37465 0
31235 0
24795 0
21030 0
15931 0
8921 0
4279 0
0 0
Mother Line
Program ID Intron size
331009 1
321859 3
315105 0
307952 32
301333 1, 4
283266 1, 4
274461 1, 36
268873 1, 15
261857 1
257954 0
249303 0
243133 1, 8, 35
233744 1, 8, 24
228057 1, 8, 24
220198 1, 24
216023 0
208013 7
202164 0
197662 0
193781 0
185935 0
177955 0
171362 0
164589 0
157003 0
149162 0
143166 0
135618 0
130161 0
123130 1, 1
88835 1, 1
80865 0
68421 0
62141 1
54857 0
42376 0
33613 0
27108 0
18502 0
7397 0
0 0

In Father Line and Mother Line, especially in Mother Line, there are several times that some introns are inherited by crossover or produced by mutation and later disappear. An obvious example is the Program 268873 in Mother Line. Its 15-node intron subtree is created in mutation from Program 261857. But when the introns disappears, I can not tell whether part of the introns are inherited by the offsprings, because the partition contributed by the parent appears in both introns and exons. The real source is unclear.


Conclusion

From the repeated experiment results shown above, the conclusion can be drawn that the Genetic Programming with introns and exons approach can increase the adaptability of the programs considerably. One noticeable thing is that the new approach does not outperform the old approach in maze 2, while it get much better results in maze 3. From this phenomenon we can go further to say that the mechanism of introns and exons affects the evolution more in complex environments.

But the detailed analysis met difficulties. The huge amount of total programs does not allow a through look of every program. And in the analysis I was able to perform, no expected phenomenon is discovered. So the possible explanation is that either the phenomenon takes place too few to see, or the introns affect the evolution in a totally different way. No matter which is the real reason, the project shows the amazing characteristic of evolution: only a small change can result in enormous difference.


Future Work

In this project I chose a toy problem for the Genetic Programming to solve. Although the population and the test numbers are large enough to get convincing results, the problem itself lacks complexity. Since my new approach seems to affect the result more in complex environment, testings on more complex problems should be performed. This is the direct future work following the project.

Another pending work is more analysis of the evolution. In this project I am not able to observe the expected phenomenon on introns because of the large amount of evolution activities. If further test can prove that this introns and exons mechanism is really able to improve the result of Genetic Programming, more concentration should be put on a micro level in the evolution to see whether some phenomena like the wings of the stick insects in [WBM03] can be observed.

Bibliography

Kam
Manolis Kamvysselis.
Evolution with a purpose.
Final paper for course 6.868, EECS, Massachusetts Institute of Technology, http://web.mit.edu/manoli/evolution/www/main.html.

Nil98
Nils J. Nilsson.
Artificial Intelligence : A New Synthesis.
Morgan Kaufmann Publishers, San Francisco, California, 1998.

WBM03
Michael F. Whiting, Sven Bradler, and Taylor Maxwell.
Loss and recovery of wings in stick insects.
Nature, 421:264-267, January 2003.

About this document ...

CPSC 601.73 Project
Evolution with Introns and Exons

This document was generated using the LaTeX2HTML translator Version 2002-2-1 (1.70)

Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.

The command line arguments were:
latex2html '-up_title=Course Projects' -up_url=../ -split 0 -local_icons report

The translation was initiated by Jie Gao on 2003-06-27

Footnotes

...
The numbers shown in Table 3 on Page [*] are generation numbers, starting from 0. So the first generation is 0, the second generation is 1, and so on.... In the following tables, the generation numbers have the same meanings.
... evolution2
In Table 5 on Page [*], the number of programs produced by evolution in all generations should be 252000 in theory. The loss of accuracy in C++ while converting floating point to integer causes one more program in each generation's elite group. The number of program in elite groups is greater than 63000 for the same reason.
next_inactive up previous
Up: Course Projects
Jie Gao 2003-06-27