An, P.E., Brown, M., Mills, D. and Harris, C.J.
High Order Instantaneous Learning Algorithms for On-line Training.
Int. Symp. on Signal Processing, Robotics And Neural Networks
Full text not available from this repository.
This paper examines several instantaneous learning rules that have been widely applied in the signal processing, adaptive control, and neurocontrol fields. These instantaneous least mean square adaptive algorithms are derived and their performance is assessed from a variety of viewpoints: instantaneous data storage, rate of parameter convergence, filtering modelling error and the computational cost of implementing the respective schemes. This work is focused on the requirement for on-line learning where the data is generally highly correlated and noisy, and the network must be able to generalise sensibly. It is shown that conventional instantaneous learning rules do not perform well under these conditions as more than one piece of information is required to update the parameter vector. Hence a set of higher order learning rules are proposed to overcome the deficiencies of the standard adaptive rules, and their performance is compared and assessed using two simple, informative examples.
Actions (login required)