Reproducibility of MOOP by setting the seed of random number generators

Hello everyone,

I have a question concerning multi objective optimization using the jMetal Framework. I have used the “NSGAIITSPRunner.java” (from the package org.uma.jmetal.runner.multiobjective and jMetal 5.1) and for adding reproducibility I added the line

JMetalRandom.getInstance().setSeed(10L);

at the beginning of the main-method. So the main-method looks like this:

public static void main(String[] args) throws IOException {
      JMetalRandom.getInstance().setSeed(10L);

      PermutationProblem<PermutationSolution<Integer>> problem;
      Algorithm<List<PermutationSolution<Integer>>> algorithm;
      CrossoverOperator<PermutationSolution<Integer>> crossover;
      MutationOperator<PermutationSolution<Integer>> mutation;
      SelectionOperator<List<PermutationSolution<Integer>>, PermutationSolution<Integer>> selection;

      problem = new MultiobjectiveTSP("/tspInstances/kroA100.tsp", "/tspInstances/kroB100.tsp");
      [...]

Furthermore, I used the data (kroA100.tsp and kroB100.tsp) which was provided with the framework. I then ran the program once and my FUN.tsv started with the line

“93328.0 132453.0”.

After I ran the program another time the FUN.tsv started with the line

“145839.0 101293.0”.

However, I was wondering why this is happening. I thought that if I set the seed to the same value both times the results would be the same, since the random values which are generated during selection/mutation/… are the same (and this worked for me for non-permutation problems like the ZD1). Can you help me clarify my misunderstanding or my error?

Kind regards

[matthieu-vergne]

It would be nice to have a general test which checks all the algorithms, and which passes only when reproducibility is confirmed everywhere (and fails by telling which algorithm is not reproducible). I opened a new issue for that.

[ajnebro]

You are right. Let me check why this happens.