When inducing a time series forecasting model there has always been the problem of defining a model that is complex enough to describe the process, yet not so complex as to promote data 'overfitting' -the socalled bias/variance trade-off. In the sphere of neural network forecast models this is commonly confronted by weight decay regularization, or by combining a complexity penalty term in the optimizing function. The correct degree of regularization, or penalty value, to implement for any

more »

... ular problem however is difficult, if not impossible, to know a priori. This chapter presents the use of multi-objective optimization techniques, specifically those of an evolutionary nature, as a potential solution to this problem. This is achieved by representing forecast model 'complexity' and 'accuracy' as two separate objectives to be optimized. In doing this one can obtain problem specific information with regards to the accuracy/complexity trade-off of any particular problem, and, given the shape of the front on a set of validation data, ascertain an appropriate operating point. Examples are provided on a forecasting problem with varying levels of noise.