home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!gatech!concert!rock!taco!ellner
- From: ellner@stat.ncsu.edu (Steve Ellner)
- Subject: Re: Training Networks on Chaotic Time Series
- Message-ID: <1992Jul23.180946.19112@ncsu.edu>
- Originator: ellner@esssse.stat.ncsu.edu
- Keywords: chaos, time series, neural nets
- Sender: news@ncsu.edu (USENET News System)
- Organization: Statistics, NCSU
- References: <670@trwacs.fp.trw.com>
- Date: Thu, 23 Jul 1992 18:09:46 GMT
- Lines: 33
-
-
- In article <670@trwacs.fp.trw.com>, erwin@trwacs.fp.trw.com (Harry Erwin)
- reminds us that neural nets performed poorly on chaotic time series
- data generated by the logistic map:
-
- |> ... the network was clearly learning the training set and not
- |> generalizing. This was particularly clear for the more highly chaotic
- |> cases since the effectiveness of the network was significantly reduced for
- |> a fixed training period. It was also clear that the network "hadn't a
- |> clue" for regions of the chaotic process that were not represented in the
- |> training set.
-
- Our experience on similar problems has been better. But, we found that
- it was essential to have a cautious criterion for adjusting the network
- complexity to match the amount and quality of training data, and
- to be _very_ persistent at seeking best-fit net parameters, i.e.,
- a stringent convergence criterion, and many, many replicate
- attempts at fitting with different starting parameters. (Refs: D. Nychka
- et al., J. Royal Statistical Soc. Series B, vol54(2), 399-426 (1992) ;
- A.R. Gallant & H. White, Neural Networks 5,129-138; and Dan McCaffrey's 1991
- Ph.D thesis, to appear this Fall in J. Amer. Stat. Assoc.). M. Casdagli
- (Physica D35, 335-356 (1989) ) also found that neural nets were among
- the best predictors for chaotic time series (see his Table I).
-
- I expect that any regression model, not just nets, will have trouble
- "learning" what's going on far from any data that are used to fit the model.
- That may not be a problem if generalization near the training data suffices
- for the task at hand. For example a short-term predictor only needs
- to be accurate near the support of the system's invariant measure (since
- real trajectories will always lie in that region) and that's where you would
- get training data from observing the system running "on its own".
-
- --- Steve Ellner (ellner@stat.ncsu.edu)
-