Examples

The following examples demonstrate the basic functionality provided by the MOEA Framework. Links to the full source code are provided alongside each code snippet. You may also find these and more examples in the demo application on the downloads page.

Setup

In order to run these examples or use the MOEA Framework, Java 8 (or a later version) must be installed on your computer. The Java 8 development kit (JDK) for Windows and Linux can be downloaded here.

To run these examples, first download and extract the latest compiled binaries from the downloads page. Windows users may extract the downloaded file using 7-zip. The files will extract to a folder called MOEAFramework-3.7. This folder will look similar to:

  • MOEAFramework-3.7/
    • docs/
    • examples/
    • javadoc/
    • lib/
    • licenses/
    • pf/
    • COPYING
    • launch-diagnostic-tool.bat
    • moeaframework.properties
    • README.md

All of the examples below are in the examples/ folder. You may compile and run an example using the following commands. Run these commands in the Command Prompt from the MOEAFramework-3.7 folder.

              javac -cp "examples;lib/*" examples/Example1.java
              java -cp "examples;lib/*" Example1

If you receive the message 'javac' is not recognized as an internal or external command, operable program or batch file, try the following steps to setup your environment on Windows or Linux. Unix/Linux users should replace the semicolons (;) with colons (:).

Example 1: Simple Run

Create and solve the bi-objective DTLZ2 test problem using NSGA-II. At the end, we display the Pareto front.


Problem problem = new DTLZ2(2);

NSGAII algorithm = new NSGAII(problem);
algorithm.run(10000);

algorithm.getResult().display();

  

Displaying the Pareto front will print all the decision variables, objectives, an constraints (if any).

Var1     Var2     Var3     Var4     Var5     Var6     Var7     Var8     Var9     Var10    Var11    Obj1     Obj2     
-------- -------- -------- -------- -------- -------- -------- -------- -------- -------- -------- -------- -------- 
0.000001 0.499800 0.499459 0.493195 0.512840 0.501684 0.496219 0.510478 0.497724 0.490473 0.498708 1.000436 0.000002 
1.000000 0.493931 0.496459 0.504963 0.492117 0.487517 0.495901 0.506916 0.498368 0.508016 0.504378 0.000000 1.000443 
0.967343 0.500695 0.493870 0.493354 0.496931 0.489186 0.496910 0.489151 0.498778 0.499109 0.496012 0.051294 0.999038 
0.570016 0.500760 0.493834 0.485232 0.492111 0.487856 0.498199 0.507383 0.498864 0.502869 0.504378 0.625569 0.780878 
...
  

Example 2: Quality Indicators

Quality indicators are used to compare results between different algorithms. Here, we calculate the hypervolume and generational distance relative to a reference set.


Problem problem = new DTLZ2(2);

NSGAII algorithm = new NSGAII(problem);
algorithm.run(10000);

NondominatedPopulation result = algorithm.getResult();

NondominatedPopulation referenceSet = PopulationIO.readReferenceSet("pf/DTLZ2.2D.pf");
Hypervolume hypervolume = new Hypervolume(problem, referenceSet);
GenerationalDistance gd = new GenerationalDistance(problem, referenceSet);

System.out.println("Hypervolume: " + hypervolume.evaluate(result));
System.out.println("GD: " + gd.evaluate(result));

  

Running this program produces the following output:

Hypervolume: 0.20834956047273037
GD: 0.001119315294859704
  

Example 3: Customizing Algorithms

The examples above used default parameters for each algorithms. But, each algorithms is customizable! In this example, we setup NSGA-II to use the Parent Centric Crossover (PCX) operator, use a population size of 250, and include an archive to store the best solutions.


Problem problem = new DTLZ2(2);
NSGAII algorithm = new NSGAII(problem);

algorithm.setInitialPopulationSize(250);
algorithm.setVariation(new PCX(5, 2));
algorithm.setArchive(new EpsilonBoxDominanceArchive(0.01));

algorithm.run(10000);

algorithm.getResult().display();

  

Example 4: Statistical Comparison of Algorithms

Larger experiments are performed using the Executor and Analyzer. The Executor is resonsible for configuring and running algorithms. Unlike the previous example where we solved one problem, the Executor is useful when running multiple seeds, multiple algorithms, or different configurations. The Analyzer performs the statistical analysis. It can compute various performance indicators, including hypervolume, generational distance, inverted generational distance, additive ε-indicator, spacing and contribution. Additionally, Kruskal-Wallis and Mann-Whitney U tests measure the statistical significance of results.


String problem = "UF1";
String[] algorithms = { "NSGAII", "GDE3", "eMOEA" };
		
//setup the experiment
Executor executor = new Executor()
    .withProblem(problem)
    .withMaxEvaluations(10000);
		
Analyzer analyzer = new Analyzer()
    .withProblem(problem)
    .includeHypervolume()
    .showStatisticalSignificance();

//run each algorithm for 50 seeds
for (String algorithm : algorithms) {
    analyzer.addAll(algorithm, 
        executor.withAlgorithm(algorithm).runSeeds(50));
}

//print the results
analyzer.display();

  

Running this script produces the output shown below. We can see that GDE3 and NSGA-II produce the best (largest) hypervolume values. Furthermore, we have determined statistically that there is no significant difference in performance between GDE3 and NSGA-II.

GDE3:
    Hypervolume: 
        Min: 0.4389785065649592
        Median: 0.4974186560778636
        Max: 0.535166930530847
        Count: 50
        Indifferent: [NSGAII]
eMOEA:
    Hypervolume: 
        Min: 0.35003766343295073
        Median: 0.47633216464734596
        Max: 0.53311360537305
        Count: 50
        Indifferent: []
NSGAII:
    Hypervolume: 
        Min: 0.38868701091987184
        Median: 0.5040946740799506
        Max: 0.5371138081508796
        Count: 50
        Indifferent: [GDE3]
  

Example 5: Collecting Runtime Dynamics

Runtime dynamics provide insight into the behavior of an optimization algorithm throughout a run. For instance, one can observe how solution quality changes with the number of function evaluations (NFE). The Instrumenter class records the runtime dynamics.


Instrumenter instrumenter = new Instrumenter()
    .withProblem("UF1")
    .withFrequency(100)
    .attachElapsedTimeCollector()
    .attachGenerationalDistanceCollector();
		
new Executor()
    .withProblem("UF1")
    .withAlgorithm("NSGAII")
    .withMaxEvaluations(10000)
    .withInstrumenter(instrumenter)
    .run();
		
instrumenter.getObservations().display();

  

The output from this script, shown below, shows how the generational distance metric changes over time. We see that NSGA-II is rapidly converging to the reference set (the optimal solutions) since its generational distance is converging to 0.

NFE   Elapsed Time GenerationalDistance 
----- ------------ -------------------- 
100   0.048455     0.554547             
200   0.068240     0.486866             
300   0.076951     0.876918             
400   0.082156     0.690796             
500   0.087247     0.542534             
...                 
9500  0.351868     0.044477             
9600  0.354273     0.037876             
9700  0.356552     0.038029             
9800  0.358757     0.038815             
9900  0.361296     0.032959             
10000 0.363735     0.020477   
  

Example 6: Defining New Problems

While we provide many test problem suites for comparing optimization algorithms, we can also introduce new problems. As demonstrated below, we need to define two methods: newSolution and evaluate. The newSolution method defines the problem representation (the number and types of its decision variables). The evaluate method takes a solution and computes its objective function values.


public class Srinivas extends AbstractProblem {

	public Srinivas() {
		super(2, 2, 2);
	}

	@Override
	public void evaluate(Solution solution) {
		double[] x = EncodingUtils.getReal(solution);
		double f1 = Math.pow(x[0] - 2.0, 2.0) + Math.pow(x[1] - 1.0, 2.0) + 2.0;
		double f2 = 9.0*x[0] - Math.pow(x[1] - 1.0, 2.0);
		double c1 = Math.pow(x[0], 2.0) + Math.pow(x[1], 2.0) - 225.0;
		double c2 = x[0] - 3.0*x[1] + 10.0;
		
		// set the objective values - these are being minimized
		solution.setObjective(0, f1);
		solution.setObjective(1, f2);
		
		// set the constraint values - we treat any non-zero value as a constraint violation!
		solution.setConstraint(0, c1 <= 0.0 ? 0.0 : c1);
		solution.setConstraint(1, c2 <= 0.0 ? 0.0 : c2);
	}

	@Override
	public Solution newSolution() {
		Solution solution = new Solution(2, 2, 2);
		
		solution.setVariable(0, new RealVariable(-20.0, 20.0));
		solution.setVariable(1, new RealVariable(-20.0, 20.0));
		
		return solution;
	}
		
}

  

Then, we can solve this problem using:


Problem problem = new Srinivas();

NSGAII algorithm = new NSGAII(problem);
algorithm.run(10000);

algorithm.getResult().display();

  

Conclusion

In addition to the above examples, we provide many more inside the examples folder. Navigate to the downloads page to download the MOEA Framework or visit our GitHub page for more information.