CGP++ is a modern and versatile C++ implementation of Cartesian Genetic Programming (CGP). CGP++ provides a generic framework that enables flexible and versatile usage. The general motivation and philosophy behind CGP++ is to facilitate the discovery of new problem domains and the reproducibility of existing results. CGP itself is a graph-based representation model for genetic programming, a search heuristic used for program synthesis. Based on object-oriented design, CGP++ provides a suitable framework that can be used to integrate new sophisticated methods that entail the implementation of complex architecture. CGP++ is not only intended to be a implementation of CGP but rather to serve as a blueprint for a modern and general genetic programming toolkit based on C++.
- Generic design
- Checkpointing
- Concurrency support
- Mutation pipelining
- Crossover support
- Parameter interfaces for hyperparameter tuning
- Benchmarking utilities
- Benchmark generator for symbolic regression
- PLU reader for logic synthesis benchmarks
Since CGP++ uses modern techniques of C++, it has to be compiled with a version of GCC that supports the C++17 standard.
g++17
The user guide provides supporting details for the usage of CGP++.
We recommend compiling CGP++ using the G++ compiler with the following command:
g++ -std=c++17 -O3
For Linux, a classical Unix makefile is provided in the build folder. The commandmake all can be used to create the executable.
For building CGP++ on Windows, we recommend, using popular choices such as:
- Mingw-w64 (https://www.mingw-w64.org/)
- MSVC (https://visualstudio.microsoft.com/vs/features/cplusplus/)
- VSCode (https://code.visualstudio.com/docs/languages/cpp)
On MacOS, the use of clang (https://clang.llvm.org/) in combination with VSCode can be considered as an alternative to using G++: https://code.visualstudio.com/docs/setup/mac
The compiled executable can be used with the parameter file in the data/parfile folder and the
following commandline options. The data file, .plu file for logic synthesis and .dat file for symbolic regression, are
located in the data/ folder. The data and parameter file are passed to CGP via commandline. Optionally, a checkpoint file
can be passed to CGP this way. The CGP and run specific options can be the passed to CGP via the commandline and parameterfile configuration.
usage: DATAFILE PARFILE CHECKPOINTFILE <options>
-a - search algorithm: 0 = one-plus-lambda; 1 = mu-plus-lambda
-n - number of function nodes
-v - number of variables
-z - number of constants
-i - number of inputs
-o - number of outputs
-f - number of functions
-r - maximum arity
-p - mutation rate
-c - crossover rate
-m - number of parents (mu)
-l - number of offspring (lambda)
-e - maximal number of fitness evaluations
-g - goal (ideal) fitness
-j - number of jobs
-s - global seed
-1 - duplication rate
-2 - max duplication depth
-3 - inversion rate
-4 - max inversion depth
algorithm - 0 = one-plus-lambda, 1 = mu-plus-lambda
levels_back - type: integer
num_jobs - type: integer
num_function_nodes - type: integer
num_variables - type: integer
num_constants - type: integer
constant_type - 0 = Koza
num_outputs - type: integer
num_functions - type: integer
max_arity - type: integer
num_parents - type: integer
num_offspring - type: integer
max_fitness_evaluations - type: integer
ideal_fitness - type: generic
minimizing_fitness - 0 = false, 1 = true
crossover_type - 0 = block, 1 = discrete
crossover_rate - type: float, range: [0.0, 1.0]
probabilistic_point_mutation - 0 = deactivated, 1 = activated
single_active_gene_mutation - 0 = deactivated, 1 = activated
inversion_mutation - 0 = deactivated, 1 = activated
duplication_mutation - 0 = deactivated, 1 = activated
point_mutation_rate - type: float, range: [0.0, 1.0]
duplication_rate - type: float, range: [0.0, 1.0]
inversion_rate - type: float, range: [0.0, 1.0]
max_duplication_depth - type: integer
max_inversion_depth - type: integer
neutral_genetic_drift - 0 = deactivated, 1 = activated
simple_report_type - 0 = deactivated, 1 = activated
print_configuration - 0 = deactivated, 1 = activated
evaluate_expression - 0 = deactivated, 1 = activated
num_eval_threads - type: integer
generate_random_seed - 0 = deactivated, 1 = activated
global_seed - type: long long
report_during_job - 0 = deactivated, 1 = activated
report_after_job - 0 = deactivated, 1 = activated
report_simple - 0 = deactivated, 1 = activated
report_interval - 0 = deactivated, 1 = activated
checkpointing - 0 = deactivated, 1 = activated
checkpoint_modulo - type: integer
write_statfile - 0 = deactivated, 1 = activated
The following call to CGP++ triggers a run with a simple 1-Bit adder benchmark by using a 1+1 ES as search algorithm and a simple CGP with 10 function nodes and point mutation:
./cgp data/plufiles/add1c.plu data/parfiles/cgp.params -a 0 -b 10 -n 10 -v 1 -z 0 -o 1 -f 4 -r 2 -m 1 -l 1 -p 0.1 -c 0.0 -e 1000000 -j 10 -g 0
Template settings:
typedef int EVALUATION_TYPE;
typedef int GENOME_TYPE;
typedef int FITNESS_TYPE;
This call triggers a run with the simple Koza1 benchmark (quartic polynomial) by using a 4+8 ES as search algorithm and discrete recombination:
./cgp data/datfiles/koza1.dat data/parfiles/cgp.params -a 1 -b 10 -n 10 -v 1 -z 0 -o 1 -f 4 -r 2 -m 4 -l 8 -p 0.1 -c 0.5 -e 10000000 -j 10 -g 0.01
Template settings:
typedef float EVALUATION_TYPE;
typedef int GENOME_TYPE;
typedef float FITNESS_TYPE;
Template types can be set in the template/templates_types.h file.
A corresponding paper that describes CGP++ in more detail is currently under review. The CGP++ team also works tirelessly on a tutorial which will be released soon.
A program output could be, for example, the overview of the configuration of the parameters and the metrics of the respective instances such as number of fitness evaluations, best fitness and runtime.
CGP++ CONFIGURATION
-------------------------------------------------
Number of function nodes: 10
Levels back: 10
Number of functions: 4
Maximum arity: 2
Number of variables: 1
Number of constants: 1
Number of inputs: 2
Number of outputs: 1
Crossover rate: 0.5
Mutation rate: 0.1
Number of parents (mu): 4
Number of offspring (lambda): 8
Ideal fitness value: 0.01
Number of jobs: 100
Maximum number of fitness evaluations: 1000000000
Maximum number of generations: 125000000
Global seed: 1707214445176910306
Functions: [1] ADD [2] SUB [3] MUL [4] DIV
Constants: [1] -0.416173
Recombination: Discrete Crossover
Mutation: [1] Probabilistic Point [2] Inversion [3] Duplication
Algorithm: mu-plus-lambda
-------------------------------------------------
Job # 1 :: Evaluations: 1980264 :: Best Fitness: 2.08206e-05 :: Runtime (s): 40.5062
Job # 2 :: Evaluations: 167528 :: Best Fitness: 2.11298e-05 :: Runtime (s): 2.47478
Job # 3 :: Evaluations: 113000 :: Best Fitness: 2.10702e-05 :: Runtime (s): 2.098
Job # 4 :: Evaluations: 49800 :: Best Fitness: 2.11298e-05 :: Runtime (s): 0.590241
Job # 5 :: Evaluations: 8832 :: Best Fitness: 2.08057e-05 :: Runtime (s): 0.128364
CGP++ is free open-source with a BSD-style academic license (AFL 3.0).
CGP++ is dedicated to Dr. Julian Francis Miller (1955 - 2022), who as the founder of Cartesian Genetic Programming devoted a large part of his scientific life to its proposal, development and analysis. With the introduction of CGP++ we pay tribute to Julian’s pioneering effort in the field of graph-based GP and acknowledge his lifework.