I use NSGA-II to solve multi-objective feed formulation with simulated binary crossover and polynomial mutation as parameters and cardinal priority ranking get best solution from non-dominated solutions. In the test case, best solution gotten when the crossover probability is 0.2 and mutation probability is 0.8. It is normal to have that values (crossover probability much smaller than mutation probability)?
Dear EKo
It has to be. In the nature also it has got slow rate of mutation. If you increase it then the will be tremendous changes in the nature and also in population.
Dear Eko
Mutation probability should very small. You need to look again at your crossover operator.
Best
Dear EKo, when we increase the mutation probability we are closer to random search locally, where we give more crossover probability to increase the exploration rate and find better solution and avoid falling in local optimal solution and try to find global optimal solutions. usually recommend more than .6 for crossover and less than .1 for mutation. best regards
Dear EKo, when we increase the mutation probability we are closer to random search locally, where we give more crossover probability to increase the exploration rate and find better solution and avoid falling in local optimal solution and try to find global optimal solutions. usually recommend more than .6 for crossover and less than .1 for mutation. best regards
Hello Dear
In discrete problems like as Graph Coloring your mutation rate should be larger than crossover and in continuous problems like as linear and nonlinear problems your crossover rate should be larger than mutation.
GoodLuck
Dear
If you are discarding the individuals from the population earlier, we will miss some of the characters which may be valuable for the next generations. So even, if GA takes time no problem but it gives better and optimum solutions than premature solution when we increase the rate mutation after some limit.
In some evolutionary algorithm, there is no crossover. For example, in many evolution strategies only an asexual reproduction take place. Human are very good at making assumptions, which may limit them to tune best the algorithm. Self-adaptation have been well studied to adapt the crossover and mutation rate to the needs to the algorithm and the problem it is attempting to solve. Some algorithm optimisation, like Grammatical Evolution has evolved algorithms with their operators as well as their mutation and crossoever. Oltean has used to do similar with with some other forms of GP.
i would recommend you to run a few experiments using a meta-evolution and then you could compare and analyse your finding much better.
Everything is based on you decisions in meta-heuristics. There aren't strict rules in this domain. If you believe and demonstrate that you obtain better results under this circumstance, then of course it is possible.
Good luck
The Crossover should have higher probability than the mutation. In your case, I think your crossover oeprator is not well designed. You should check it carefully.
Dear Patricia, thank you for you suggestion.
Dear Fehmi, yeah before, I also think that. But I doubt it because from lecturing class I join into, it explain what Zhenzhen suggest. So I try change cardinal priority ranking with fitness evaluation method to get best solution, and it solved.
Dear Eko
Generally, the crossover rate is kept in between 0.7 to 0.9 and mutation rate at 0.01 to 0.001 as to get more and more optimal solution too. But some time, the rate of mutation is kept on changing as and when the solution goes towards the saturation.
Usually in Genetic Algorithms (GA), the crossover /rate/probability is kept higher (0.6-0.9) and mutation rate/probability is kept far smaller (close to zero ... 0.001, 0.005, 0.01 etc.). The higher crossover rate helps you explore the solution space more (and gives your algorithm global search capability). The lower mutation rate helps you explore local vicinity of your solutions (giving you local search capability). However, there is no constitutional ban on changing these probabilities to meet your needs. If different rates/probabilities work for your specific problem and then go for it (the only thing is that it would not remain a GA in the traditional sense). It may be a consequence of the some special structural characteristics of your problem (I am not a student of the problem domain you mentioned!).
[BTW, I would also suggest you to check your own cognitive definition as well as software implementation of these operators. It may well be the case that some software routines are switched in implementation. May be what you are calling a GA is more or less an MA (Memetic Algorithm) where you almost always resort to local improvement (some very extensive local search) in the solution generated by the traditional GA component of the algorithm)]
In fact, if you do not find some literature in your problem domain supporting such probabilities then some quick simulation results showing "convergence rates" and "quality of result" etc for various combinations of probabilities would actually be a good research publication in itself. i.e. just reporting this would be a very good contribution to the literature (it would useful to others in your domain to work better, as they would be able to choose more 'helpful' mutation and crossover rates). You may also gain some new insights into the structure of the problem and may end up formulating the problem in a novel and improved manner. It would give more "novelty" (discovery) to your research paper and may help you easily publish in higher ranking journals (just beware that calling it a GA may not be considered appropriate by the reviewers ... so, exploring some appropriate name may be in order).
Good luck.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages