Output details

Advanced adaptive nonmonotone conjugate gradient training algorithm for recurrent neural networks

<24> Although humans learn in a nonmonotone way, deterministic training algorithms mainly focus on monotone approaches. We challenge this view by developing nonmonotone methods that build on a strong mathematical basis. Training performance is better than other mainstream approaches using FFTD, LRN and NARX Recurrent NNs, in different types of problems. This is a substantially extended version of a paper presented at the 19th IEEE ICTAI 2007 (not returned to RAE 2008), which was recommended for submission to IJAIT. The journal version presents a new algorithm that improves the performance, and also provides a theoretical justification for the method, demonstrating applicability to various cases.