Actually the method we use could be classified as reinforcement learning. Most (?) evolutionary algorithms can use the concept of reinforcement learning to explore the state space of the model (if I remember correctly this was included in Hollandâs work on GA). We have looked at SVM but havenât tried it. The hope is that kernel regression could reduce computation times. I have no experience of RVM. The total size our network(s) is similar to yours. Our model also includes a number of other parameters not included in the NN. /Hugin
Donât want to turn this thread into a GA specific thread, but here we go. We started with population sizes recommended by Grefenstette (i.e. pretty small ~30) but found that we had to increase it a bit, possibly due to the fact that our problem differs from standard benchmark problems. We also found papers by Goldberg by that pointed to a population size of ~60. This estimate changes a little depending of what type of model weâre optimizing but a size around 80 seems to be working well. Regarding the number of generations we set it fairly high (>100) and break a run when the population has not improved over the last 10 generations, as measured by change in average fitness and the diversity of the population. We also have a stop criteria if the elite portion of the population has not changed in the last 10 generations. /Hugin
There are better ways to do it. Just Google "Markov Genetic Algo", you'll find papers regarding them...
Thanks, will do if time permits. The Grefenstette study, even if old, is interesting since they use a meta level GA to optimize GA parameters over a number of benchmark problems. The Goldberg study is based on schema theory. /Hugin
Did a quick search and found a few articles. Even found a few on island/migration GA. From these it seems that the Markov chain analysis is mostly directed towards population convergence (most stipulate population to be âlarge enoughâ or use the population size as an input to the analysis) and not how to determine a poulation size, but maybe I haven't found the right papers. Another thing is that we do not use binary string genomes which raises a question. How will this impact the transition matrix in the Markov chain analysis, especially with respect to mutation? Anyway, I found some good pointers on convergence that might become useful. /Hugin