Today in our Friday Thursday Paper Seminar we’ve been discussing the paper «Fuzzy System Parameters Discovery by Bacterial Evolutionary Algorithm» of Norberto Eiji Nawa and Takeshi Furuhashi.
It’s a very interesting paper because introduces a bioinspired algorithm based in bacteria (as the title says, obviously), and its application to fuzzy system. The main improvement from the previous algorithm of the authors the Pseudo-Bacterial Evolutionary Algorithm is the fact of using the transduction, the way that the bacterias interchanges information between themselves. But let’s begin from the beginning.
Every individual of the population is a bacteria whose chromosome is the parameters of a fuzzy system. This chromosome has low epistasis, that is, the chromosoma have weak interrelationships between parts so it is possible to perform optimization in parts. So in each generation every bacteria have a local improvement called bacterial mutation, cloning the bacteria and improving every part but taking care of the global fitness, so the best clone is chosen.
The second operator presented is the Gene Transfer Operation. Instead crossover, the best bacterias sends information to the worst bacterias adding or overwriting parts of the chromosome.
After explain the algorithm the authors performs several experiments to compare other algorithms and this one. I’m not expert in Fuzzy Systems, so probably I’ll be wrong, but I can see that the while the BEA has lower error rates evaluating train set, it also has higher error rates in test set, and the fuzzy system obtained has more (and not used) fuzzy rules. I think this is not a good improvement, but the idea of bacterial mutation and gene transfer is quite interesting in other problems, as we can see in more papers that refers this one.
So, that’s all! Until next week!
Pablo (aka Fergu)
GeNeura Team 