Chapter Evolving Connectionist and Fuzzy Connectionist Systems: Theory and Applications for Adaptive, On-line Intelligent Systems

12. Problems with ECOS and 
EFuNNs and directions for
further research
In spite of the advantages of ECOS and 
EFuNNs when applied for on-line,
adaptive learning, there are some difficulties that should be addressed in the
future. These include finding the optimal values for the evolving parameters, such
as the sensitivity threshold Sthr, the error threshold Ethr, learning rate lr1 and lr2,
forgetting rate. Also, pruning of fuzzy rule needs to be made specific for every
application, thus depending on the definition of age and the other fuzzy variables
in the pruning rule. One way to overcome the above problems is to regularly apply
genetic algorithms and evolutionary computing as optimisation procedures to the
ECOS and EFuNN structures. Introducing DNA computing as part of the evolving
process may also be beneficial [65].
Evolving connecti onist systems could be considered as a 
new AI paradigm.
They incorporate the following AI features: learning; reasoning; knowledge
manipulation; knowledge acquisition ; adaptation. This sets new tasks for the
ECOS future development that are relevant to the current AI methods and
systems, such as implementing in ECOS non-monotonic and temporal reasoning,
optimal dimensionality spatial representation, 
meta-knowledge and 
meta-
reasoning. More theoretical investigations on the limitations of ECOS are needed.
13. Conclusions
This chapter presents a framework ECOS for evolving connectionist and fuzzy
connectionist systems, and evolving fuzzy neural networks EFuNN, in particular,
for building on-line, adaptive learning systems. Real problems have been used to
illustrate the potential of this approach. Several applications of ECOS and
EFuNNs have been demonstrated in the chapter, namely: adaptive speech
recognition; financial applications; 
multi-modal information processing and
integrated audio and video information processing; intelligent agents. ECOS have
features that address the seven major requirements to the next generation 
neuro-
fuzzy techniques presented in section one.

---

140