Although genetic algorithms perform globalised search for optimal solutions, they are computationally expensive and sensitive to the selection of various parameters such as population size, crossover and mutation probabilities. Have you considered using other machine learning techniques such as inductive learning which does not suffer from the above mentioned problems? Have you compared your new method with other techniques currently available?

Thank you for your interest and for your questions. In every engineering design problem there is a space instantiated by the alternative solutions.  It is quite common for these solution spaces to be highly non-linear, with discrete solution intervals, while the available design knowledge could be vague and expressed with formalisms other than analytical relations. Considering the above, we decided that the optimal design solution must by searched by using a non-gradient based heuristic algorithm, that could avoid local extremes in the solution space, while having the possibility of balancing between speed and robustness. Several techniques have been tested in the problem of oscillating conveyors’ design. To mention a few of them: ant colonization, simulated annealing, tabu search, memetic algorithms, pattern search etc. The results show that GAs were classified somewhere in the average. Their adjustability in trading-off speed with accuracy and the significant raise of their performance if used in combination with a traditional optimization technique, led as to their deployment. The GA is deployed only for optimization purposes and they do not participate actively in the learning process.
As far as the learning process is concerned, there are many alternative approaches that can be deployed. It is true that Inductive Logic Programming could deliver a design model suitable for re-design (despite the low learning rate). Also other approaches based on hybrid neural networks could be also used. The neuro-fuzzy learning presented in the current work dominated because we considered the following two issues:
a. the fuzzy rules, delivered by the trained ANN, are extracted from elite solutions b. they can be further calibrated rationally in a future re-design.c) the fuzzy rules that model the simplified design model can also be used in a collaborative framework among designers with aggregation of the existing fuzzy sets under a specific aggregation strategy.  
Our approach is a part of an innovative research work in the context of an integrated domain-independent parametric design framework and can only be compared to existing approaches if separate modules are considered each time.
I hope that my answers are sufficient enough.  

Could you please give values of all parameters (such as population size, mutation rate, etc.) that you used? Thanks.

interesting paper to read.but i'am not quite clear about "evolution of the problem under consideration" as one of 4 basic guidelines of your work, which mentioned in the second paragraph on page 3 of pdf version. can you please make clear what it does mean and how your new methodology achieved it?

thanks.