home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!wupost!cs.utexas.edu!qt.cs.utexas.edu!yale.edu!ira.uka.de!Germany.EU.net!infko!inews
- From: evol@infko.uni-koblenz.de
- Subject: Technical report on topology optimization of anns available
- Message-ID: <1993Jan07.061204.9289@infko.uucp>
- Sender: inews@infko.uucp (inews)
- Organization: University of Koblenz, Germany
- Date: Thu, 07 Jan 1993 06:12:04 GMT
- Lines: 34
-
- Our new Technical Report "Synthesis and Performance Analysis of Multilayer
- Neural Network Architectures" is now available via ftp:
-
-
- ftp archive.cis.ohio-state.edu or ftp 128.146.8.52
- cd pub/neuroprose
- binary
- get schiff.gann.ps.Z
- quit
-
- In this paper we present various approaches for automatic topology-optimization
- of backpropagation networks. First of all, we review the basics of genetic
- algorithms which are our essential tool for a topology search. Then we give a
- survey of backprop and the topological properties of feedforward networks. We
- report on pioneer work in the filed of topology--optimization. Our first
- approach was based on evolutions strategies which used only mutation to change
- the parent's topologies. Now, we found a way to extend this approach by an
- crossover operator which is essential to all genetic search methods.
- In contrast to competing approaches it allows that two parent networks with
- different number of units can mate and produce a (valid) child network, which
- inherits genes from both of the parents. We applied our genetic algorithm to a
- medical classification problem which is extremly difficult to solve. The
- performance with respect to the training set and a test set of pattern samples
- was compared to fixed network topologies. Our results confirm that the topology
- optimization makes sense, because the generated networks outperform the fixed
- topologies and reach classification performances near optimum.
-
-
-
- Randolf Werner
-
-
-
-
-