Backpropagation BP and Related Networks

The operation of a BP network has been fully described elsewhere (see in particular Chapter 8 of Ref. 19), so we merely summarize its operation here. During the learning process, input variables flow toward the output layer, while 

In many cases it is difficult to determine in advance how many hidden layers and how many HL PEs are required for satisfactory performance. A trial-and-error method to determine this information can be very time-consuming. Cascade correlation networks build HLs one PE at a time, solving a problem incrementally.  

BP networks tend to get stuck in local minima. Although the algorithms mentioned earlier that speed up learning help to overcome this tendency, it may be best to use a different network if you think this is a serious problem. The networks listed in the second to last paragraph for regression and classification problems are possible options. 

Nuclear magnetic resonance (NMR) has also been a fruitful area of ANN BP application. 3-in Most studies have dealt with either the simulation of spectra or the prediction of shifts, although one study focused on prediction of phosphorus shifts. A BP network was used to predict secondary protein structure, which was then used to assist in NMR assignment. In another study, H NMR spectra of binary mixtures of alditols were successfully classified.  

BP networks were used for multivariate calibrations of pyrolysis mass spectra. 2. 3 In one case the BP networks gave better concentration predic- 

---