Use Genetic Programming To Generate File Converters

Introduction

Here is an old article I wrote in January 2009. Since then, I received some emails and a lot of 404 errors, so I decided to repost it on this “new” blog. You can find the code here on github.

Genetic algorithm, another metaheuristic inspired by evolution of species. You can read the description of the algorithm on the genetic algorithm article on Wikipedia.

I applied the algorithm to a concrete problem: help lazy programmers and non-programmers to write file converters. I had to write file converters to unify all (more or less) formatted input files into one kind of CSV file. Each of these converters is made using the right combination of filtering / regexp matching / line splitting.

Description

So I wrote a program based on genetic algorithm where an individual is composed by 3 “genes”:

  • a list of filters (remove consecutive spaces, replace tabs by spaces, …)
  • a list of basics regexp to extract formatted basic blocks(int, float, date, line separator, …)
  • a list of cleaning functions (remove empty cell, merge cells, …)

The fitness of an individual is calculated using a sample file to parse that goes through the genes (filters / regexp / cleaning functions). The result of this process is a CSV file that can be compared to the expected sample result (that I wrote manually). The comparison uses the SequenceMatcher of the difflib Python stantard module, it returns the fitness: a float. A fitness close to 1.0 means the results is close to the expected CSV format. When the fitness is exactly 1.0, that means the converter works perfectly for the sample.

Here is a sample file I wanted to convert:

08/12        Hello world 06/12
Paris 121231231         - 22,29
08/12   Something something ... 04/12
1111 1111 1111 1111         - 14,35
08/12   something else
        - 12,96
26/11   Vir AAAAA
AAAAAA 2008     264,51

into this csv file:

"08/12","Hello world 06/12","Paris 121231231","-22,29"
"08/12","Something something ... 04/12","1111 1111 1111 1111","-14,35"
"08/12","something else","","-12,96"
"26/11","Vir AAAAA","AAAAAA 2008","264,51"

The program prints for each generation the 5 bests individual (I’m using elitism, so the best individual is always kept from a generation to the next one). A sample run:

$ python ga.py tests/sample1.txt tests/sample1.expected result1.pickled
Generation: 0 (mutation rate=10)
0.53772
0.42507
0.30406
0.28598
0.26389
Generation: 1 (mutation rate=10)
0.53772
0.42507
0.32684
0.30406
0.26799

...

Generation: 271 (mutation rate=15)
1.0
0.96025
0.95107
0.91779
0.87886
"08/12","Hello world 06/12","Paris 121231231","-22,29"
"08/12","Something something ... 04/12","1111 1111 1111 1111","-14,35"
"08/12","something else","","-12,96"
"26/11","Vir AAAAA","AAAAAA 2008","264,51"

filters: str_remove_consecutive_spaces, str_remove_somespaces, str_strip, str_remove_consecutive_spaces
regex: ([0-9]{2}/[0-9]{2})(.+?)(
)(.+?)(-?[0-9]+[.,]+[0-9]+)
cleaners: clean_strip, _in, _in, _in

Conclusion

The good:

  • It’s fun to watch your program evolve ;)
  • You are lazy and don’t want to write _many_ of this kind of file converters. This was used in a real life problem, with more than 200 different input file formats.
  • With a good interface and if the basic functions and genes size are well-defined, a non-programmer can create his own parser.

The bad:

  • The resulted parsers are not optimal

The ugly:

  • Of course it’s quite easy to write such genes by hand
  • It may never converge to 1.0 and generated parsers of fitness != 1.0 are useless (this may not be the case for other kind of GA applications)
  • You need many well-chosen primitives to cover a wide set of solutions

What next? Write a simple Web interface to help people generate their own converters. Add new construction blocks. Use the same idea to automatically plot any kind of formatted data.