Evoptool: Evolutionary Optimization Tool

From AIRWiki
(Redirected from Evoptool)
Jump to: navigation, search


Evolutionary Optimization Tool (Evoptool) is an open source optimization tool writte in C++, distributed under GNU General Public License. Evoptool implements meta-heuristics based on the Evolutionary Computation paradigm and aims to provide a common platform for the development and test of new algorithms, in order to facilitate the performance comparison activity. Evoptool offers a wide set of benchmark problems, from classical toy samples to more complex tasks, and a collection of algorithm implementations from Genetic Algorithms and Estimation of Distribution Algorithms paradigms. Evoptool is flexible, easy to extend, also with algorithms based on other approaches other from EAs.

The Evoptool Team

Formerly involved

  • Emanuele Corsano

PPSN 2012

In this section we report the details of the experiments presented in the paper "Variable Transformations in Estimation of Distribution Algorithms", submitted to PPSN 2012.

The experiments were run using the unstable version of evoptool, which is available trought SVN at


For each problem, Alternated Bits, Trap3, Trap3 Overlapping and Trep5, a python script is used to generate the Evoptool configuration files. For a given problm size n, we chose a population size, run 24 instances of the test algorithm and check the success ratio. If it below 95%, we increase the population, otherwise we compute from these runs the average number of fitness function evaluations performed before the global optimum appears in the population. We increase the population exponentially with n, so that N=n^k and each time the success ratio is below the chosen we set k=k+0.1.

Each time an experiment is performed using Evoptool, a .tar file is produced which contain statistics and plots for each instance of the algorithms run and also aggregated averages. The python scripts we used name this tar files with the following convention:


All the .tar files used to compute the result presented in the paper, along with the python scripts and everything else needed to reproduce the experiments are available trought SVN at


Mailing List



Algorithms implemented

  • Genetic Algorithms
    • SGA
  • Estimation of Distribution Algorithms
  • DEUM framework
  • Stochastic Gradient Descent
    • SGD, SNGD
  • Other Meta-heuristics
    • Simulated Annealing

References to original papers will be added soon.

Benchmark functions implemented

  • OneMax
  • AltBit
  • Trap3
  • Trap5
  • IsingSpinGlass2D (*)
  • IsingSpinGlass3D (*)
  • Max Cut

(*) instances generated and solved with Spin Glass Server. This service is provided by COPhy and M. Jünger's group. http://www.informatik.uni-koeln.de/spinglass/


 svn checkout https://svn.ws.dei.polimi.it/evoptool/trunk --username usrevoptool --password evoptoolpsw

Developers may also access unstable branch. Contact the evoptool team for more details.


Follow these steps in order to compile evoptool

1. Download source code from svn

 svn checkout https://svn.ws.dei.polimi.it/evoptool/trunk --username usrevoptool --password evoptoolpsw

Refer to trunk for the lastest (stable?) version of evoptool

2. First you need to manually compile the l1_logreg-0.8.2 package, http://www.stanford.edu/~boyd/l1_logreg/ whose source are already included in the evoptool repository. In order to do that follow the instrunctions in README.evoptool in the l1_logreg-0.8.2 directory

Next, compile <id_dist/code> following the instructions in <code>README.evoptool in the <id_dist folder

3. Then you need to compile a module at a time. Each module is included in a different folder. To compile a module, from go to the module source folder and do make lib. For instance, for the module named common

 cd common/src
 make lib
 cd ..

Compile modules in the following orderm due to dependencies


To clean a module, do

 make clean

from the module source directory

4. Go to core module and type

 make exe

Binary files will be copied in the bin directory in the root as well as the bin directory of the core module

Required packages

  • Software
    • gnuplot
    • R (lars package)


Right now evoptool only runs as a script. GUI is not maintained in the latest version. The evoptool binary file is supposed to be run in the bin directory in the root, this is because right now the script looks for some configuration files, for the instances of the problems, and temp directories. If you want to run the script from other directories, you just need to copy in that directory the evoptool-file directory that you find in the bin directory. Link such directory instead of copying it is not safe if you run multiple scripts at the same time, since output files will be shared.

To run the algorithm you need to enter as input an xml file. For instance you can run

 ./evoptool evoptool-file/examples/unitTestOneMax.xml

The evoptool-file/example directory contains a set of xml files for different benchmarks and different algorithms. Take a loot at the example.xml file for the documentation on hoe to set the parameters for an each execution of evoptool

Each execution of evoptool produces a set of files as output. You can find all files in evoptool-file/temp directory. Plus a tar.gz file containing such files can be produced at the end of the execution of evoptool. See the xml file for details.


contains the raw data of the statistics according to the xml file


contains the gnuplot files to produce images of the statistics


contains images of the statistics produced according to the xml file


contains logs


Implement a new benchmark function

Benchmarks are defined in the function module. A benchmark is the maximization problem of a real valued function defined over a vector of n binary variables. Functions are maximized in evoptool. In order to implement a new benchmark function you have to create a new class that inherits from ObjectiveFunction in the common module. Take a look at the OneMax benchmark function.

Among the arguments of the constructor of the ObjectiveFunction class there must be the size of the benchmark, which may or may not be a parameter of the new benchmark.

 OneMax::OneMax(int size) : ObjectiveFunction(size)

In the constructor you have whether the maximum of the function is known or not

  _knownSolution = true;

And in case this is known, set the minimum value and maximum value for the function

_minFitness = 0; 
_maxFitness = size; //the maximum value of the OneMax function is defined as the sum of the bits of x set to one.

Besides, such values are used in the normalization of the functions in the plots of the statistics.

Each benchmark must implement the f method, which takes the BinaryString x and return f(x). For instance, for OneMax you have

 double OneMax::f(BinaryString *bi) {
  if (bi != NULL) {
   if(!(bi->validCache())) { 
    /* Fitness cache value not valid */
    int sum = 0;
    for (int j = 0; j < _size; j++) {
     sum = sum + bi->get(j);
    return sum;
   } else {
    return bi->getFitnessCache();
  } else {
   cerr << "[OneMax::f] binary instance cannot be null" << endl;
   return 0;

In order to avoid multiple evaluations of the same BinaryString x, a caching mechanism has been implemented.

Moreover, you are asked to implement the exportInteration function, which return an HyperGraph which represents the interactions present in the function. In order to determine such hypergraph, start from the polinomial representation of the function, and for each monomial introduce a new hyperedge in the hypergraph.

For the OneMax function, such representation corresponds to the independence graph, since such function is linear.

 HyperGraph* OneMax::exportInteractions() {
  HyperGraph *interactions = new HyperGraph();
  return interactions;

In Evoptool.h add an entry in the Tasks enum.

 /* This enumeration defines an identifier for each task function. */
 typedef enum {
  ONE_MAX = 7,
 } Tasks;

This value will identify uniquely the benchmark in the xml file. In the XMLParser.cpp file, add an entry for the new function.

 bool XMLParser::parseTask(xmlpp::TextReader &reader, Evoptool::TestParameters *params) {
  switch (task) {
   case Evoptool::ONE_MAX: params->task = new OneMax(size);
   default: return false; 

Additional parameters can be obtained using the xml TextReader object, as for the ForPeaks function.

 case Evoptool::FOUR_PEAKS: 
  value = readNodeContent(reader, "peaks", &errorFlag);
  if (errorFlag) return false;
  int peaks = toInt(value);
  params->task = new FourPeaks(size, peaks);

Finally in the xml file, a function can set with the task tag, together with all parameters.


To create a new algorithm

1. Implement a new algorithm

All algorithms must inherit from the Algorithm class. First override the Algorithm::initAlgorithm as follows:

 void DEUM::initAlgorithm(ObjectiveFunction* of) {
   /* Initialization tasks */

Then implement a run() method which performs al the required task for one iteration of your algorithm.

There are some classes which contain useful methods such as samplers and estimators from which your algorithm can inherit. The most useful are:


2. Append entry in the Algorithms enum in Evoptool.h.

 typedef enum {
   DEUM    // 5
 } Algorithms;

In the xml configuration file, your algorithm will be characterized by the int value associated with the position in this enum declaration.

3. modify the file core/src/XMLParser.cpp. Search for the method bool XMLParser::parseAlgorithms and insert an new entry for your algorithm in the algorithmParameters array, for example

 AlgorithmParameters algorithmParameters[] = {
   { Evoptool::DEUM, 5, (char*) "IDDDI" }, // Pop, percElitism, percSelec, CoolRate, GibbsIterations

In each entry, the first field is the name of the entry in the Algorithms enum just modified in Evoptool.h, the second is the number of parameters to be parsed in the xml configuration file and the third is a string specifying their type: I for integer and D for double.

4. add a case in the switch that follows in bool XMLParser::parseAlgorithms, for instance:

 case Evoptool::DEUM: 
   algos[i] = new extendedFCA(intVal[0], doubleVal[0], doubleVal[1], intVal[1], intVal[2], params->rng);

Note that your algorithm can employ the random number generator provided by the Evoptool core, params->rng.

5. Include your headers in core/src/Algorithms.h

6. Document how your algorithm is instantiated in the comments appearing in evoptool-file/exec-scripts/example.xml

  Algorithm  Id    Types    Parameters
   FCA       26    "IDDII"  (Pop, PercSelection, LearningRate, MaxCompositionLenght, MaxMonomialOrder)

7. Add an instantiation example for your algorithm in evoptool-file/exec-scripts/example.xml

 <algo id="26" instanceName="FCA">  

Run an algorithm in a C++ program

The simplest way to do this is to create a new submodule in core, i.e., a new source directory within the source directory of a module For instance, the steps to create a new submodule named myexample are

1. copy the example submodule exampleRunAlgorithm in a new folder named myexample in


2. modify the Makefile in the myexample directory, see comments in the makefile for more details.

3. edit the main function in the source file, if you need you can create other C++ and .h files in the same folder.

4. from evoptool/core/src/myexample run

 make bin

The bin file is create in evoptool/core/bin and copied in evoptool/bin

NOTICE that evoptool/core/bin may be deleted after a make cleanall

5. from evoptool/bin run


Available statistics

In the configuration xml file you can enable the evaluation of three set of statistics:


See the headers file in evoptool/common/src for details.