2

I have written a uniform crossover algorithm for part of my homework but it's not working properly. It's actually returning worse results than my one point cross over. I would just like someone to point out where I am going wrong so I can fix it please :). I've been trying for ages now and this is my last resort!!

    private void DoUniformCrossOver(int p1id,int p2id)
{
    ArrayList<Integer> p1 = population.get(p1id).GetRep();
    ArrayList<Integer> p2 = population.get(p2id).GetRep();
    ArrayList<Integer> c1 = new ArrayList<Integer>();
    ArrayList<Integer> c2 = new ArrayList<Integer>();

for (int i=0;i<nbits;++i)
{
    double selected = CS2004.UI(1,2);
    if (selected ==1)
    {
        c1.add(p1.get(i));
        c2.add(p2.get(i));
    }
    else
    {
        c1.add(p2.get(i));
        c2.add(p1.get(i));
    }
}

    population.add(new ScalesChrome(c1));
    population.add(new ScalesChrome(c2));
}

The method takes in as paramaters the two parents, p1id and p2id. Then creates the arraylists of the representation - p1 and p2.

In the for loop, 'nbits' is the weight of the array (or the length of the array). My one-point crossover method uses it in the for loop and it works just fine.

I then generate either 1/2 to determine which gene from each parent the child will get.

The fitness of this algorithm is very very poor!! Any help at all would be greatly appreciated.

Many thanks.

manlio
  • 18,345
  • 14
  • 76
  • 126
Tamara JQ
  • 359
  • 2
  • 7
  • 14
  • So you are generating two lists, that have elements from either the first or second source list. The algorithm seems to do just that, i don't see anything wrong with it matching what it's trying to do. So the problem i guess is your algorithm. What are you trying to accomplish, and/or why is this performing poorly in your estimation. – MeBigFatGuy Mar 19 '11 at 19:41
  • +1 for your clear, well-prepared question. – Pete Wilson Mar 19 '11 at 19:44
  • Yeah.. It's generating two lists (children c1 and c2) which are a combination of their parents (p1 and p2). There is a 50% chance that a child will get the first gene from parent 1 and a 50% chance it will get it from parent 2... I was told to implement the algorithm and that it is more effective than the one point crossover but my one point crossover performs brilliantly and this doesn't!! :-S! – Tamara JQ Mar 19 '11 at 19:45
  • explain why this is less brilliant? When you modify only one item, you will have a better chance of progressing to your goal, as that has more predictable hill climbing characteristics, however, the one point crossover method will never escape local maxima in the solution domain, this is where the above algorithm will do better. – MeBigFatGuy Mar 19 '11 at 19:47
  • Thanks @Pete. Well @MeBigFatGuy I've created a fitness function, the closer to 1 the fitness function is, the better. The one-point crossover returns values of 1 almost every time, whereas this one returns 13 25 39 3 etc... There are much larger values generated in the output! – Tamara JQ Mar 19 '11 at 19:51
  • 1) Are you sure of the randomness of `CS2004.UI(1,2)`? 2) you're comparing a double with a constant 1.0 which is problematic for floating points numbers, using `int selected` will probably do better. – rsp Mar 19 '11 at 20:47
  • Did you find a solution ? is the uniform mode better ? – slim Oct 30 '22 at 20:22

2 Answers2

1

Well, first of all what kind of information are you encoding and what are you trying to evolve?

Depending on the problem you are trying to solve, some kinds of cross-over strategies will prevent you from ever finding good solutions.

A simple example: if the solution you are looking for has inherent symmetry (e.g. a white/black cellular automata majority classifier), single point cross-over will never give you very good results because it's breaking any symmetry that the genetic algorithm may have stumbled upon (e.g. so in the case of the majority classifier, it will be very good at classifying black or white but not both, so it will never get better than a given - pretty low - fitness).

JohnIdol
  • 48,899
  • 61
  • 158
  • 242
0

(As you didn't respond to my comment, I'll repeat it as answer)

You're comparing a double with a constant which is problematic for floating points numbers, using int selected will probably do better, or in case you need to use a double, try using something like:

if (selected < 1.5)
rsp
  • 23,135
  • 6
  • 55
  • 69
  • Hey @rsp, sorry i wouldn't have ignored you, i didn't see a comment from you until now. I've tried what you've said but no success! I'm at the end of my tether with it now! I have a lab on Tuesday so I will ask someone then! Thanks very much though :) – Tamara JQ Mar 20 '11 at 12:21
  • @Tamara, did you add debug logging to see what the results of this method are? Do you get random mixes of the parents? (Unexpected results of the fact that you always create mirror-image children maybe?) – rsp Mar 20 '11 at 12:53
  • Hiya,I'm not familar with debug logging? Sorry, but I've done the print outs for the children c1 and c2, and the parents p1 and p2 and the algorithm seems to be working correctly in that way. I was just under the impression it was supposed to return a better fitness result, but it isn't =(. – Tamara JQ Mar 20 '11 at 15:51