This project contains code to optimize the PacMan genes by an evolution algorithm. Note that this code is a Python library rather than a stand-alone program. So it can be used in a very flexible way inside any Python script.
Since this code contains native C modules for efficiency reasons in making a large number of duels, the code needs building to work. To build the native C modules, on the top-level directory of this project, the following command can be run
python3 setup.py build
Since there are OpenMP parallelization over the duels in a tournament, the
compiler is required to support the -fopenmp compiler flag. Then by running
python3 setup.py install, the code can be installed. Then a simple Python
code like,
import evolpacshould work.
All the code are in the evolpac package. Inside it, the subpackage duel
contains the core functions for running duels among genes. The dueling
mechanism is treated as black box outside this package. Basically it contains
two native modules, duel.c contains code to run a duel between two given
genes, and tournament.c runs duels among all genes in a pool in parallel and
gets the average score for each gene. The functions in the native modules are
all forwarded in the duel subpackage. So just importing the subpackage is
sufficient for using the facilities.
In the top-level evolpac package, the module gene_manip contains some
utility functions for the manipulation of the genes. evolve is the core
module contains the core function for running the evolution. Note that this
contains a very general function supporting all kinds of tweak of the parameters
in the evolution algorithm. And we can also output the intermediate results
into JSON files. The drivers module was intended for large high-level driver
functions. It actually just contains a driver to generate a given number of
genes from repeated short contest among randomly generated genes. The module
vizres is a separate module for the visualization of results. For instance,
it can be used to visualize the content in result.json by
python3 -m evolpac.vizres result.json
then we can get a sneak peek into the result like
Digest Gene Weight Score Eval
327c47 03000000110100003330321121231211221231131131100133 1 17.165 13.278
50f1f7 03000000311100003300321121221211221231132130300131 14 18.780 12.556
d2b2b4 03000300111100003300321121221211221231131131330110 2 11.209 15.111
816a36 03000000110100003330321121211211221231131131123311 1 16.810 11.444
1cf4fd 03000000110100003330321121211211221231131130130131 10 10.891 15.611
89bd06 03000300111100003300321121221211221231131131103131 1 17.596 14.056
193c87 03000000110100003330321121211211221231321133130330 8 16.551 10.778
a49d8f 03000000111100003330321121221211221231131130301131 1 16.140 14.000
d7a5bf 03000000110100003300321121221211221231131133303130 1 15.705 13.167
9f6d8f 03000000100100000333321121231211221231131130130131 20 15.978 17.278
eec437 03000000110100003300321121221211221231101133130130 2 14.275 16.000
c0024f 03000000311100003300321121221211221231131130300131 6 10.825 15.444
where the digest are the SHA-1 hash of the gene sequence, which can be used for easier naming and tracking of the genes.
Due to the low quality of really randomly generated genes, we would first like a pool of genes with relatively good quality. Then later random generation of genes could just pull a random gene from this pool. This can be achieved by the following Python code.
from evolpac.drivers import form_pool
with open('pool.txt', 'w') as out_fp:
form_pool(3000, out_fp, pop_size=600, evol_steps=100)which will attempt to generate a pool of 3000 genes, each of which are the winners among a random population 600 genes evolved by 100 steps. Next, we can run a real long evolution by the following code,
from evolpac.duel import run_tournament, eval_w_sample
from evolpac.evolve import evolve
from evolpac.genemanip import gen_random_gene, get_gene_from_str, read_gene_strs
def main():
"""The main driver."""
with open('../pool.txt', 'r') as fp:
pool = read_gene_strs(fp)
good_ones = [
get_gene_from_str('03100003010300103333321121121213221121131133132300'),
get_gene_from_str('03000001111122201313211212111212212211211311101301'),
get_gene_from_str('03000001111122201303211212111212212211211310001331'),
get_gene_from_str('00000020111122200300111111211211221221330330310333'),
get_gene_from_str('03000000032123203301211232212211212211133131103033'),
get_gene_from_str('01310000110020203333131323323322222322323310323310'),
get_gene_from_str('01310000110020203330131323323322222322323313023310'),
get_gene_from_str('03000001111122201313211212111212212211211311311300'),
get_gene_from_str('03000001111122201303211212111212212211211311311300'),
get_gene_from_str('03000000111022203333211112212211212211311231021301'),
get_gene_from_str('03000000111022203333211112212231212211032123322331'),
get_gene_from_str('03000000110100003330321121221211221231131133101131'),
get_gene_from_str('03000300011100003330321121221211221231131130221110'),
get_gene_from_str('03000000011100003330321121221211221231131133301131'),
get_gene_from_str('03000000000100003330321121221211221231131133330131'),
get_gene_from_str('03000300000100003330321121221211221231131133301131'),
get_gene_from_str('03000300010100003330321121221211221231131133130131'),
get_gene_from_str('03000000011100003330321121221111221231131133131131'),
]
eval_sample = good_ones
evolve(600, 10000, run_tournament, init=good_ones, eval_cb=eval_w_sample(eval_sample),
gene_db=pool, out_steps=100)
if __name__ == '__main__':
main()In this script, we start with a few good genes that we have found in previous evolutions, the rest of the population are all drawn from the pool that we just generated. Then this population is evolved by the evolution algorithm.