STUDIES IN BIOLOGICALLY INSPIRED COMPUTING

Article type: Research Article

Authors: Esat, I. I.

Affiliations: School of Engineering and Design, Brunel University, Uxbridge, UK

Abstract: This paper presents some studies in biologically inspired computing. The case studies are grouped under two sections, in the first section case studies involving more fundamental aspects of networks and in the second section application oriented studies are presented. The first section involves geometrical interpretation of back propagation networks and a scheme for generating network structure from the decision space was offered. The second study presents a new encoding scheme and also studies how the encoding scheme and performance are related. The third study is specific to Hopfield neural networks and shows that a stiffness matrix has properties which allow it being mapped to the Hopfileld transition matrix. The last study presents the particle swarming algorithm and a new concept of quantum inspired computing. The second section involves cases relating to some engineering applications or bench marks of some biologically inspired computing. The case studies as many of them based on manipulation of engineering structures, this section presents a stiffness matrix of a simple spring attached on a 3D body. The following study involves manipulation of matrix elements in order to achieve vibration decoupling (of modes). This is studied in three different ways; firstly an analytical solution is obtained. This way, the equations were proved to be ill conditioned. The second study involves using these equations with GA algorithm which showed that GA is robust in dealing with ill conditioned problems and finally it was assumed no prior knowledge of the matrix formulation. In this case GA operated directly with the mode shapes obtained from the eigenvector analysis. The final approach appears to be the most unreliable in convergence. The matrix formulation was also used with Hopfield networks. The other examples included gearbox fault detection which as well as identifying the faults the study demonstrated the importance of preprocessing and information filtering or refinement. Furthermore the studies on gear box faults showed that even the type of preprocessing influences the performance of identification. The study of fault identification in gas cylinders led recognition of various loopholes in manually selection of features for fault identification. The study involved manual feature extraction which in this case led to endless experiments, however the use of the frequency response function (FRF) proved to be the most effective in this case and probably it is likely to be the same for similar cases where the system disturbance levels cannot be repeated. The last case study involved bench mark based comparative study among GA, PSO and quantum inspired computing. This study showed that all three are similar in their effectiveness but quantum inspired computing algorithm showed an incredibly fast convergence during the first few cycles of optimization. A series of less detailed cases such as cellular automata for breast cancer growth modeling, the use of GA in real time control, were also included in the paper. The paper, finally weighs findings against the No Free Lunch (NLF) theorem, although conclusions are complimentary. The work presented illustrates the reliance of biologically inspired computing to representation and encoding schemes both for architecture and data for its effective execution.

Keywords: ANN design, GA, PSO, biologically inspired, quantum inspired, stochastic search

Journal: Journal of Integrated Design & Process Science, vol. 11, no. 3, pp. 85-104, 2007