SmcGP_deap

DEAP implementation of Sub-machine Code GP

A presentation of the project can be found in the /docs folder.

DEAP homepage and sources at https://pypi.org/project/deap/

To install the latest version: pip install git+https://github.com/DEAP/deap@master

DEAP documentation at https://deap.readthedocs.io/en/master/

Needed packages

DEAP

see [deap installation guide] (https://deap.readthedocs.io/en/master/installation.html).

Graphics

To plot the tree `pygraphviz' is needed. It has some dependecies (assuming python3)

sudo apt-get install python3-dev graphviz libgraphviz-dev pkg-config
sudo pip install pygraphviz

Datasets

Datasets of digits taken from license plate images are provided, in which each line represents a different pattern along with the corresponding digit/label (last number in each row). The patterns are encoded as:

• four 32-bit words (as in "": scanning the patterns row-wise, the first 3 words encode lines 1-3, 4-6, and 7-9 of the input pattern, respectively, into the least significant 24 bits; the fourth 32-bit word encodes the last 4 rows of(10-13) of the input pattern) bigall32c.csv and bigalltest32c.csv

• two 64-bit words (scanning the pattern row-wise, the first word encodes lines 1-6 of the input pattern into its least significant 48 bits, the second word encodes lines 7-13 into its least significant 56 bits). bigall32c_64_int.csv and bigalltest32c_64_int.csv

• 104 bits (13 rows x 8 columns of the input pattern scanned row-wise) These files may be helpful to examine/plot the patterns with no need to convert the integers back into bits. bigall32c_104.csv and bigalltest32c_104.csv

Run training and test

Smc_GP_main.py: this script performs the following

1. Initilizes a dictionary which holds the parameters.

2. Runs Smc_GP_test.py, which trains the classifier.

3. Runs eval_best_tree.py, which runs the best classifier on the test set and dumps the results.

Each of these scripts is autonomous and data are passed between them using .pkl files.

NOTES

The file eaSimple_cp.py contains a modified version of eaSimple, which dumps periodic checkpoints of the evolution.

############################################################## NB The scripts provided work ONLY with 64-bit encoded data!!!! ##############################################################

To use 32-bit data it should be enough (not guaranteed) to modify the files Smc_GP_training.py, fitness_fun.py, and eval_best_tree.py as follows:

• change 'word_len' into 32 in Smc_GP_main.py

• change the fitness function prototype to require 4 parameters pset = gp.PrimitiveSetTyped("MAIN", [int, int, int, int], int) in Smc_GP_training.py and eval_best_tree.py

• change the input file names in Smc_GP_training.py and eval_best_tree.py

• read 4 integers per pattern from the input files instead of 2 for j in range(0, row_num): ARG0 = bigall3_matrix[j][0] ARG1 = bigall3_matrix[j][1] ARG2 = bigall3_matrix[j][2] ARG3 = bigall3_matrix[j][3] label = bigall3_matrix[j][4] in Smc_GP_training.py and fitness_fun.py