Information Technology Reference
In-Depth Information
Machine Learning Techniques
With such as large number of forecasting techniques and parameters available for tuning them, it be-
comes more difficult to choose the appropriate one in a particular context. In addition, as the complexity
of patterns present in the noisy data increases, such as demand at the end of supply chain, it becomes
difficult to choose an appropriate model for forecasting purposes. One possible solution is to rely on a
class of algorithms called “universal approximators”, which can approximate any function to an arbi-
trary accuracy. Using such universal approximators, any function within the in the time series data can
be learned. This effectively makes traditional forecasting techniques a subset of the functions that the
universal approximators can learn. Machine learning (ML) techniques, such as artificial neural networks
(ANN) and support vector machines (SVM) are examples of universal approximators.
Forecasting time-series such as those in supply chains involves a data domain that is very noisy. It is
highly desirable only to learn true patterns in the data that will be repeated in the future and to ignore
the noise (e.g. random error). The, ML-based techniques have two important features that are useful for
supply chain forecasting problems in the presence of noise: (i) the ability to learn an arbitrary function
and (ii) the ability to control the learning process itself.
Artificial neural networks (ANN) and recurrent neural networks (RNN) are frequently used to
predict time series data (Dorffner, 1996; Giles, Lawrence, & Tsoi, 2001; Herbrich, Keilbach, Graepel,
Bollmann-Sdorra, & Obermayer, 1999; Landt, 1997). For example, because manufacturer's demand in
general can be a chaotic time-series, RNNs can be trained using so-called “backpropagation of error”
through time, which permits them to learn patterns through to an arbitrary depth in time. Support vector
machines (SVM), a more recent learning algorithm that has been developed from statistical learning
theory (Vapnik, 1995; Vapnik, Golowich, & Smola, 1997), have a strong mathematical foundation and
have been previously applied to time series analysis (Mukherjee, Osuna, & Girosi, 1997).
neural networks
Neural networks have been used successfully in the past for forecasting in complex business domains.
Some examples of such applications include predicting foreign exchange rates (Walczak, 2001) and
expert judgments in bankruptcy prediction (Kim & McLeod, 1999). The most commonly used artificial
neural networks are the “feed-forward-error, backpropagation” type. In these networks, the individual
elements (“neurons”) are organized into layers in such a way that output signals from the neurons of a
given layer are passed to all of the neurons of the next layer. Thus, the flow of neural activations goes
in one direction only, layer-by-layer (feed-forward). Errors made by the neural network are then used
to adjust all the network weights by moving back through the network (error backpropagation). The
smallest number of layers is two, namely the input and output layers. More layers, called hidden lay-
ers, could be added between the input and the output layer and the non-linear transfer function of the
hidden layers is to increase the computational power of the neural nets. ANNs have been proven to be
universal approximators assuming that sufficient hidden layer neurons are provided and assuming that
the activation function is bounded and non-constant (de Figueiredo, 1980).
Search WWH ::
Custom Search