A set of libraries and tools to create and test Genetic Algorithms experiments for combinatorial optimization problems.
With the javascript library and the React.js GUI it is possible to create and test Genetic Algorithms experiments with focus on hyperparametric optimization for solving the most common problems present in the combinatorial optimization literature.
The C++ implementation, on the other hand, is available here, faster than the javascript one, but it not as flexible as the javascript implementation.
This project was developed under the context of the final work for the posgraduate course "Advanced Techniques for Evolutionary Computation" by Dr. Ignacio Ponzoni at DCIC (UNS).
I hope the reader find this project helpful. I have not much knowledge in software architecture or design patterns, but my intention was to make readable code and a useful tool for those who are teaching or learning about Genetic Algorithms or Numerical Optimization Techniques.
The main contribution of the library was written in javascript. This allows you to share the code through any medium and because its easy for someone to get access to an internet browser, almost every user will be able to run this code, on almost any smart device. Consider this as the only and most important advantage of this software, against other tools as Python or Matlab.
Lets solve the Subset Sum Problem using this library. Here I'm using Node.js, but with a few adjustments, you can make it work on the browser too.
import Ga from 'optimization/ga/index.mjs';
import SubsetSum from 'optimization/fitness/subsetsum.mjs';
// We're using the following numeric set of 20 elements
const set = [-96, -91, -87, -84, -82, -75, -71, -27, 12, 30, 46, 53, 73, 79, 80, 88, 90, 94, 94, 95];
const target = 0;
// We create the already implemented fitness model
const f = new SubsetSum(set, target);
// And the GA optimizer attached to this fitness model, configuring the mutation probability as 5%.
const ga = new Ga(f, {mut_prob: 0.05});
// Then, we run 100 generations
for(let gen = 0; gen < 100; gen++)
ga.evolve();
// Solution is in the first chromosome, as the population list is always sorted from best to worst
const solution = ga.population[0];
// Finally, we ṕrint results
process.stdout.write("Best subset: "+f.decode(solution.genotype)+"\n");
process.stdout.write("Objective value: "+solution.objective+"\n");And the output will be something like:
Best subset: -87,-82,-75,30,46,80,88
Objective value: S = 0, N = 7
Try the latest version here or use this application locally running the following commands (Node.js already installed is required):
$ git clone https://github.com/matiasmicheletto/dna-solver.git
$ cd dna-solver
$ npm install
$ npm run build
$ npm run startIf you need to use just the optimization module via scripting (without GUI), checkout the examples folder and run the scripts installing the optimization package only (and cli-progress for this example, but not necessary), using the following commands:
$ git clone https://github.com/matiasmicheletto/dna-solver.git
$ cd dna-solver
$ npm install cli-progress ./optimization
$ node examples/tsp/example_tsp_selection.mjsThis library provides a class to model any objective function with an interface to be optimized using Genetic Algorithms. Five class examples are provided to show how to extend this class in order to model common combinatorial optimization problems. The Ga class implements a Genetic Algorithm optimizer with many configuration options (see next section). Finally, the Experiment class allows to create and run different experiments to test the behaviour of GA optimizers when configuring different hyperparameters.
To create a new Fitness model, extend the prototype class, for example:
export default class MyNewFitness extends Fitness {
constructor(param1 = 1, param2 = 3) {
// First we need to call the constructor of the parent class,
// and pass the attributes or parameters:
super({
_param1: param1,
_param2: param2,
_name:"My new fitness model"
});
// Then we can use this._param1 or this._param2 as we need.
}
objective(x) {
// This example just implements a simple linear function:
return x * this._param1 + this._param2;
}
decode(g) {
// Suppose we're using 16 bit BCD to decimal conversion.
return parseInt(g.join("").slice(-16), 2);
}
objective_str(g) {
// This function shows the result of evaluating the objective function
// as a human-readable string.
return "F("+g+")="+this.objective(this.decode(g));
}
eval(g) {
// This is the fitness function. This function should return a numeric scalar
// value that represents the solution's quality.
return this.objective(this.decode(g));
}
rand_encoded() {
// As we're using binary strings, then the random solution generator will
// return a random binary array with 16 bit length length:
return new Array(16).fill(0).map(() => Math.round(Math.random()));
}
get ga_config() {
// Lets say you don't want the user to know how to properly configure
// the GA method to use your fitness model, so we can facilitate a
// default configuration:
return {
pop_size: 50
mut_prob: 0.01,
cross_prob: 0.1,
selection: selection.TOURNAMENT, // Remember to import "selection" from "ga"
mutation: mutation.BITFLIP, // Remember to import "mutation" from "ga"
tourn_k: 4 // As we're using tournament, we set the tournament size to 4
};
}
}And thats it, now we can make our first experiment to see how does this behave (spoiler: will behave pretty bad, as its just a linear function):
import Experiment from 'optimization/experiment/index.mjs';
import MyNewFitness from 'mynewfitness.mjs'
const experiment = new Experiment(); // Create the experiment manager
const f_id = experiment.add_fitness(MyNewFitness, [2, 8]); // Add our fitness with some parameters
experiment.add_ga(f_id); // Attach an optimizer to our fitness
// Run the experiment!
experiment.run({
rounds:100,
iters:25,
progressCallback:p => process.stdout.write("Progress = "+p+"% \n")
});
// Ptint results:
process.stdout.write(experiment.getPlainResults());The following table shows the configuration parameters and default values used by the "Ga" class module to implement GA optimization.
| Parameter | Type | Default value | Description |
|---|---|---|---|
pop_size |
Integer number greater than 4 | 20 |
Population size, number of chromosomes |
elitism |
Integer number between 0 and pop_size |
2 |
Number of elite individuals. Elite individuals are force-preserved through generations |
selection |
ROULETTE, RANK or TOURNAMENT |
ROULETTE |
Selection operator enumerator |
crossover |
SINGLE, DOUBLE, CYCLE or PMX |
SINGLE |
Crossover operator enumerator |
mutation |
BITFLIP, SWAP or RAND |
BITFLIP |
Mutation operator enumerator |
cross_prob |
Float number between 0 and 1 | 0.5 |
Crossover probability (probability that a pair of selected individuals to be crossovered) |
mut_prob |
Float number between 0 and 1 | 0.1 |
Mutation probability (probability of an gen to change). Usually 1/(bitstring length) |
rank_r |
Float number between 0 and 2/(pop_size*(pop_size-1)) |
0.002 |
Ranking parameter (In case of ranking based selection). High r increases selective pressure |
tourn_k |
Integer number between 2 and pop_size |
3 |
K parameter for tournament selection method. Usually between 2 and 5 |
best_fsw_factor |
Float number between 0 and 1 | 0.2 |
Window size for getting evolution slope value proportional to generation number |
param_control_enabled |
Boolean | false |
Enable or disable the automatic parameter control |
controlled_param |
CROSS_PROB, MUT_PROB, RANK_R or TOURN_K |
CROSS_PROB |
The controlled hyperparameter |
param_control_factor |
Number | 0.01 |
The incremental factor of the controller parameter |
controller_var |
GENERATION, POP_S2, EVOL_SLOPE or POP_AVG |
GENERATION |
The controller variable (static control by default) |
The last four parameters are used in automatic parameter control. There are two operation modes, static or adaptive. For static control, then GENERATION should be used as controller variable, then the controlled parameter will increment its value in factor (positive or negative) units on every generation until it reaches its maximum or minimum value. Otherwise, in the case of adaptive control, the controlled parameter will increase or decrease its value in factor*value units, where value is the numeric value of the controller variable, which can be EVOL_SLOPE (evolution slope), POP_S2 (population variance) or POP_AVG (population average fitness).
A React.js and Bootstrap GUI allows to build experiments graphically. There are two components that depend on the Fitness models, "FitnessConfig" and "SolutionViewer". If not appropiate components are provided to configurate the model, then the FitnessConfig section will be displayed as a blank or empty space, and the SolutionViewer will show the solution vectors as dash-separated-element strings. Some ReactJS knowledge is required to code and include the components for a new Fitness model, but the ones provided will be helpful to understand the idea.
This tool is the most recent and there is still work to be done, especially refactoring the implementations in order to improve readability, modularity and maintainability. It fulfils a basic function and gives the users the possibility to easily modify its internal structure for adapting it to their needs.
This example shows how to solve the Subset Sum Problem.
// Import other required libraries
#include <iostream>
#include <vector>
#include <random>
#include "ga.h" // Adjust the path to the ga.h file
// We're using the following numeric set of 20 elements
std::vector<int> set = {-96, -91, -87, -84, -82, -75, -71, -27, 12, 30, 46, 53, 73, 79, 80, 88, 90, 94, 94, 95};
// We define the gene model, which consists of a binary string of 20 bits.
class BoolGene : public Gene {
public:
BoolGene() : Gene() {
randomize();
}
inline void randomize() override{
digit = rand() % 2;
}
inline void print() const override {
std::cout << digit << " ";
}
inline bool getValue() const {
return digit;
}
inline void setValue(bool value) {
digit = value;
}
private:
bool digit;
};
// Then, we define the chromosome model. This chromosome will have 20 genes, each one representing a number in the set, where 1 means the number is selected and 0 means it is not.
class BinaryStringCh : public Chromosome {
public:
// The constructor of the parent class Chromosome requires as an argument the number of genes, because the mutation probability is set as 1/(number of genes).
// The initialization of the gene vector is carried out in the child class constructor. In this case, we create a set of the previoulsy defined genes.
BinaryStringCh(std::vector<unsigned int> *set) : Chromosome(set->size()) {
this->set = set;
unsigned int size = set->size();
for (unsigned int i = 0; i < size; i++) {
BoolGene *ig = new BoolGene();
genes.push_back(ig);
}
}
// The chromosome name is defined here and used mostly for debuggin purposes.
std::string getName() const override {
return "Subset selection array";
}
// The phenotype is information that is represented by the genes. In this case, each chromosome defines a sum value, depending on its genes values.
unsigned int getPhenotype() const {
unsigned int sum = 0;
for (unsigned int i = 0; i < genes.size(); i++) {
BoolGene *gene = (BoolGene*) genes[i];
if (gene->getValue()) {
sum += set->at(i);
}
}
return sum;
}
// The genotype is the representation of the phenotype, the print method allows to see what is the internal value of the chromosome, for example, when printing results.
void printGenotype() const override {
std::cout << "Genotype: ";
for (Gene* gene : genes) {
gene->print();
}
std::cout << std::endl;
}
// Printing the phenotype of a chromosome is useful to understand the solution found by the algorithm.
void printPhenotype() const override {
std::cout << "Phenotype: Subset = ";
for (unsigned int i = 0; i < genes.size(); i++) {
BoolGene *gene = (BoolGene*) genes[i];
if (gene->getValue()) {
std::cout << set->at(i) << " ";
}
}
std::cout << "- Sum = " << getPhenotype() << std::endl;
}
// The clone method allows to buld another chromosome with the same genes values.
void clone(const Chromosome* other) {
std::vector<Gene*> otherGenes = other->getGenes();
// To access the child class methods, we need to cast the genes
std::vector<Gene*> thisGenes = getGenes();
for (unsigned int i = 0; i < otherGenes.size(); i++) {
BoolGene *thisGene = dynamic_cast<BoolGene*>(thisGenes[i]);
BoolGene *otherGene = dynamic_cast<BoolGene*>(otherGenes[i]);
if (thisGene && otherGene) {
thisGene->setValue(otherGene->getValue());
} else {
std::cerr << "Gene cast failed" << std::endl;
}
}
fitness = other->fitness;
}
private:
std::vector<unsigned int> *set;
};
// Finally, we define the fitness model. This model will evaluate the chromosome, giving a higher value to the best solutions.
class SubSetSumFitness : public Fitness {
public:
SubSetSumFitness(std::vector<unsigned int> *set, long int target) : Fitness() {
this->set = set;
this->target = target;
}
// The name of the fitness model is defined here.
std::string getName() const override {
return "Quadratic function";
}
// The following function is the core of the fitness model. It evaluates the chromosome and returns a numeric value that represents the quality of the solution. As it can be seen, the result is computed as:
// 100 / (abs(error) + 1) - sizeCost
// where error is the difference between the target value and the sum of the selected numbers, and sizeCost is the proportion of selected numbers in the chromosome.
// This minimizes the error between the subset sum and the target value, and at the same time, it tries to minimize the number of selected numbers.
double evaluate(const Chromosome *chromosome) const override {
BinaryStringCh *c = (BinaryStringCh*) chromosome;
unsigned int subSetSize = 0;
for(unsigned int i = 0; i < set->size(); i++){
BoolGene *gene = (BoolGene*) c->getGenes()[i];
if(gene->getValue()){
subSetSize++;
}
}
const long int error = (long int)c->getPhenotype() - (long int)target;
const double sizeCost = (double)subSetSize/(double)set->size();
return abs(100.0 / ((double) abs(error) + 1.0) - sizeCost);
}
// The chromosome generator is defined here. It returns a new chromosome with random genes values. This method is defined here because the fitness model knows the chromosome structure, and the genetic algorithm need to generate new chromosomes at the initialization process.
BinaryStringCh* generateChromosome() const override {
BinaryStringCh *ch = new BinaryStringCh(set);
ch->fitness = evaluate(ch);
return ch;
}
private:
std::vector<unsigned int> *set;
unsigned int target;
};
int main(int argc, char **argv) {
// We create the already implemented fitness model
SubSetSumFitness f(&set, 0);
GeneticAlgorithm ga(&f, 0.05);
// The genetic algorithm requires the fitness function to construct the object.
GeneticAlgorithm *ga = new GeneticAlgorithm(f);
// The library provides a method to configure the genetic algorithm. This method receives the command line arguments and sets the parameters of the genetic algorithm.
// Check the manual.txt file to see the available parameters.
// There is an overloaded setConfig function that takes a GAConfig object with all the configuration parameters. If the setConfig function is not called, the genetic algorithm will use the default parameters.
ga->setConfig(argc, argv);
// The print function displays the configuration of the genetic algorithm.
ga->print();
// The run function executes the genetic algorithm and returns the best solution found.
GAResults results = ga->run();
// The print function displays the results of the genetic algorithm.
results.print();
delete ga;
return 0;
}Author: Matías Micheletto - matias.micheletto@uns.edu.ar
ICIC - CONICET - UNS
License: GPL-3.0