One of the “short-term” approaches that the Center exploits for the prediction of properties is Neural Networks.  Researchers in CAMM have used both Bayesian Neural Networks and other types of neural networks (including so-called ‘fuzzy-logic’ neural networks) to make blind and semi-blind predictions of properties.  In addition, researchers have used trained Bayesian Neural Networks in a variety of novel ways to determine functional dependencies, guide further experimental efforts, and inform researchers of regions of databases that are data-lean.
           

In general, neural network modeling is a sophisticate method of non linear regression that avoids problems with the linear regression Neural networks are more flexible, able to fit the data better and capture model regularities which are not possible with other regression models. Neural networks are “trained” to deliver the desired result by an iterative process whereby the weights associated with each input at each node are adjusted to optimize the desired output. Different transfer functions can be used to capture interaction between variables; tangent hyperbolic functions are used because of their flexibility. The strength and shape of the hyperbolic tangent can be varied by adjusting the weight wj. The degree of non linearity can be altered by number of hyperbolic tangent functions and enable to capture any arbitrary relationship between variables. Advantages of using hyperbolic tangent functions includes the ability to capture complex relationships between input and output variables and the ability to capture interactions across the entire input space.