Simulated Annealing in combination with the test functions can lead to an endless loop

[mkommend]

The problem can be reproduces by executing the attached algorithm a few times. The file contains a simulated annealing algorithm which optimizes the rastrigin function and uses the QuadraticDiscreteDoubleValueModifier (all other parameters were left at the default values).

[mkommend]

It looks like the infinite loop is a result of an invalid individual (out of bounds) in combination with the StochasticNormalMultiMoveGenerator line 60-62.

 do {
    strength = N.NextDouble();
 } while (vector[index] + strength < min || vector[index] + strength > max);

The probability that the strength has a value to correct the individual (moves it into the specified bounds) is almost 0 and this code block loops forever.

The problem is that I don't know how and where the invalid individual is created and how it surpasses the bounds checker.

[abeham]

I looked at the engine and think the reason lies in the InnerIterations of the simulated annealing. If you test with InnerIterations = 1 this bug should not occur.

This is similar to the issue addressed in #1185.

[abeham]

r9085: Added 3.4 version of SA (including a reheating strategy and limited to one move per iteration)

[abeham]

pulling from next release

[abeham]

will be fixed with #1836