What are artificial neural networks?

Artificial neural networks are tools that simulate the behaivor of mammalian neurons as they process information and learn about patterns in data. Artificial neural networks, or neural nets, belong to a class of machine learning tools within the br o ader field of artificial intelligence. Neural nets, because they are pattern recognition tools, differ from other quantitative modeling tools such as deterministic models and statistical models. They share more in common with statistical models than deterministic models.

Neural net models contain several features. First, most neural net models contain input, hidden and output vectors (Figure 1). Input vectors contain information that is quantified with output vectors through a process called "training". The training process presents data to the neural network so that values are passed through a nonlinear function (called the activation function), which is commonly a logistic or tanh function, combined together at nodes at the next layer by a squashing function (which is generally a linear combination of weights and values from the previous layer - this is shown in Figure 2 below). A bias function can be added to adjust the activiation function across the Cartesian plane.

Figure 1

click on image to enlarge


Figure 2

click on image to enlarge

Neural nets begin training by assigning randow weights between nodes in the input layer and nodes in the hidden layer and between nodes in the hidden layer and the nodes of the output layer. As information is passed forward through the neural network, the value of the output is compared to the observed value for the output as it proceeds back (called back propagation).


Figure 3 illustrates the process of feed forward of weights and back propagation and errors, respectively. As the net cycles through this process, a learning algorithm, often called the delta function, is used to determine the difference of errors between the last two cycles. If the error has decreased, then the weights are changed slightly. The errors often decrease over time, as illustrated in Figure 4 below.

Figure 3


Figure 4

click on the image to enlarge


One of the modeling exercises that is conducted during modeling is the analysis of the pattern of error over the number of training cycles (sometimes called epochs). We find that most LTM applications find a minimum error very rapidaly, sometimes after only 100 epochs. We have examined how model different node weights and activation values affect model performance. This research is important as many of the models that are coupled in this project rely on the LTM for accurate predictions.


Why Use Neural Nets?

Neural nets are powerful generalization tools. Models developed with one set of data are likely to perform well on another set that are presented in the same manner (e.g., same variables, samve neural net topology). We have begun to examine how this ability to generalize can be used to expand our modeling across space, time and datasets. For example, we have tested whether a neural net model that is built to simulate land use change in the Detroit area can be used to predict historical changes in the Twin Cities metropolitan area, and visa versa. In addition, we are testing whether small training sets can be used to scale up to larger regions. We call these exercises "spatial learning exercises".

We have also used a set of time series data for the Twin Cities to determine temporal generalization as well (Pijanowski et al., 2005). This work shows that there is nearly a 90% match between time steps across a 13 year period suggesting that the neural net model can generalize across time as well.