I think Ant Colony Optimization is one of the state of the art bio-inspired algorithms. Also it has shown a brilliant performance in both continuous and discrete search spaces.
Elephants playing chess. My AI professor's snide remark on bio-inspired algorithms is that there's one based on everything. Eventually, there will be one based on elephants playing chess.
I'm all for getting inspiration wherever it comes from, but a lot of bio-inspired stuff I've read lately spends more time talking about what is being mimicked as opposed to how these features are relevant to the task at hand.
I'd like to see more research that identifies a specific issue, and finds some inspiration (natural-based or otherwise) to address it.
Recently bio inspired algorithm have been attract many researchers to improve their algorithms. I'm currently combining some bio-inspired algo such bat, cuckoo with Back propagation and it performed much much better....
Thank you all for your answers. Right now I am looking for a good modelling bio-inspired algorithm. I have extensively used neural networks, but I believe that a change will be good. I have complex chemical problems where phenomenological models are difficult to determine or have high errors and consequently I am looking for alternatives.
@Bahram Zaeri , Indeed swarm intelligence has a rapid development nowadays.
It's not a particular problem. I just feel that when obtaining 5-10% average percent errors for a specific model, there is still place for improvement. I usually work with search spaces having both continuous and discrete parameters, multiple local minima. Also the data gathered for modelling the process has measurement errors.
Some early work on autonomous vehicles had a neural net observe a human driver to learn how to drive. When they analyzed the neural net, it was essentially doing edge detection of the side of the road. Subsequent work focused on developing better edge-detection algorithms.
General techniques like neural nets can be really useful to get a first solution, but their generality means they rarely produce the best overall solution. If you look at the solution your neural net produces, is there any insight into what type of solution it is trying to build? You might get a better solution by eliminating the neural net and just trying to build that solution directly.
In particular, your measurement errors are a major problem. If your neural net just averages them out, that's going to put a limitation on the accuracy of its solution. If you can come up with explicit models, you might have a better chance of getting past this noise.