CogLang - Genetic Algorithm

(Skip here for how to try it out.)

Use a genetic algorithm to generate words for a "Cognate Language" (a.k.a. "CogLang" for short). The "CogLang" is a language with special properties (details here). The genetic algorithm puts the words in output.txt. Results seem to resemble those generated manually or deterministically (deterministic_output.txt comes from https://github.com/hchiam/cognateLanguage).

Goals at a Glance:

Generate words:

• with as much "cognacy" as possible with the source words from all 5 source languages (like recognizability -- an emphasis on initial syllables in words should help),

• while allowing for "allophones" (acceptable alternative spellings for flexibility),

• while having the generated words relatively easy to pronounce,

and

• while having words as short as possible.

See the code in geneticAlgo_just1_v2.py for details on how the above goals are quantified and combined for overall "fitness" scores.

Background / Other Related Repos:

https://github.com/hchiam/cognateLanguage

https://github.com/hchiam/cognateLanguage2

https://github.com/hchiam/coglang-translator

Try it out:

Use the command line interface.

To run an example on just one word:

python geneticAlgo_just1_v2.py

To run on all the data in data.txt: (This outputs to output.txt and best-scorers.txt)

python geneticAlgo.py

You can compare scores by running this:

python evaluator_old.py output.txt

or this:

python evaluator_old.py deterministic_output.txt

or even to compare with a previous run with an older genetic algorithm:

python evaluator_old.py previous_output.txt

Lessons Learned:

• Find more efficient data structures and algorithms = faster and less risk of over-heating. Example: hashtable.

• Better-defined fitness evaluators and better generators = better results / more sensible results.

Stats:

Scores Using evaluator_old.py:

According to the deterministic/manual method's own rough evaluator:

output.txt 13476.4 = rough overall score

deterministic_output.txt 12201.3 = rough overall score

So:

Using the genetic algorithm is 1.105 times as good as doing it deterministically/manually, or about 11 % better, even if you use the deterministic method's own rough evaluator.