simonof@aplcen.apl.jhu.edu (Simonoff Robert 301 540 1864) (09/21/90)
Hi, I am working on a problem where I must be able to predict the outage of a system early enough so that humans can interveve and prevent the outage. Earlier I asked about using Kohonen networks as novelty detectors (as per Kohonen's book), and was not able to get any answers on how self organizing feature maps can be used as novelty detectors - this was my first approach to solving the outage prediction problem. I would like to ask the question in reference to back propagation - how would one go about solving the following type of problem using bp: You have data representing the system state. There is a lot of variables that can be fed into the network, but it is unknown exactly at what point the data represents a state which is stable and at what point the data represents a state which will lead to a system outage. My first thought was to consider the last recorded system state as an unstable state and train the network to descren this is outage eminant. But what of the state data 2 seconds earlier, is that bad or not?? (no one knows for sure). Is the only answer to assume the last recorded system state is outage predicting (say output of 0.0), use analog outputs and if the network output < 0.3 or 0.4, consider the system in trouble?? The question is how might one go about this using a supervised network? Thanks, Bob Simonoff simonof@aplcen.apl.jhu.edu (301) 240-3168
mehra@aquinas.csl.uiuc.edu (Pankaj Mehra) (09/21/90)
In article <6610@aplcen.apl.jhu.edu> simonof@aplcen.apl.jhu.edu (Simonoff Robert 301 540 1864) writes: > You have data representing the system state. There is a > lot of variables that can be fed into the network, but > it is unknown exactly at what point the data represents > a state which is stable and at what point the data > represents a state which will lead to a system outage. This problem should be solvable using the family of detection algorithms described by Sutton ("Learning to Predict ..", Machine Learning, v. 3, 1988). Also, I am surprised that you are ignoring the simplest and most well-studied techniques of linear prediction (FIR and IIR filters). The following references should help you get started: Hamming, "Digital Filters", 3 ed., 1989, Prentice-Hall if you know your undergrad. math, you can read this book in two days Widrow and Stearns, "Adaptive Signal Processing", 1985, Prentice-Hall examples of adaptive recursive filter design Goodwin and Sin, "Adaptive Filtering, Prediction, and Control",??,Prentice-Hall more advanced text describes stochastic predicition, Kalman filters -Pankaj {Mehra@cs.uiuc.edu} -- Pankaj Mehra e-mail: {mehra@cs., mehra@aquinas.csl., p-mehra@}uiuc.edu phone: (217)244-7176