NEURAL NETWORKS AND MACHINE LEARNING
If mechanisms for prediction are not well understood and historical data
is available, predictive statistical models such as neural networks may
provide a solution.
If mechanisms for prediction are not well understood and historical data
is available, predictive statistical models such as neural networks may
provide a solution.
PREDICTIVE MODELS
Predictive models such as neural networks can help to increase the quality of the product and the efficiency of the production process. Some example applications are:
FORWARD MODELING
DATA DRIVEN MODELS, NEURAL NETWORKS
The underlying functional form of the mapping is often unknown. For most problems, e.g. the ones from the applications described above, it is infeasible to create a model from first principles by hand. Fortunately, if a representative data set of input-output examples is available, then machine learning, including logistic and linear regression from conventional statistics and neural networks can be applied to create a model.
The main issue is how these models generalize with new inputs. Success relies on the availability of informative input variables and a sufficiently representative training set.
Predictive models such as neural networks can help to increase the quality of the product and the efficiency of the production process. Some example applications are:
- Prediction of sales or prediction of consumer behavior can be used to optimize production to match the demand or to optimize market strategy.
- In the making-industry, prediction of the quality of a product or of produced materials on the basis of e.g. acoustic signals can be used for an automated quality control system. This may improve both quality and efficiency.
- Prediction of physical processes. For example, for shipping guidance, where current prediction plays a major role in determining whether a deep draught ship can be admitted to a port. Similarly, the current and temperature of a river close to a power station determine the amount of waste water and thus the capacity of the power station.
FORWARD MODELING
- Expected sales as function of previous sales, season and parameters in marketing activity.
- Physical state of a material as a function of certain features in the response signal of an acoustic test.
DATA DRIVEN MODELS, NEURAL NETWORKS
The underlying functional form of the mapping is often unknown. For most problems, e.g. the ones from the applications described above, it is infeasible to create a model from first principles by hand. Fortunately, if a representative data set of input-output examples is available, then machine learning, including logistic and linear regression from conventional statistics and neural networks can be applied to create a model.
The main issue is how these models generalize with new inputs. Success relies on the availability of informative input variables and a sufficiently representative training set.
jed (Just Enough Delivery)
An example of a system that is based on predictive modeling is JED. We have developed this system for and in collaboration with the Dutch newspaper publisher De Telegraaf for the prediction of single-copy sales of daily newspapers and magazines at individual outlets.
Read more about this product
realizing applications
SMART Research BV has the tools and the expertise to build predictive (e.g. neural network) applications.
More about realizing applications
netpack
For our neural networks, we have developed a neural network library in MATLAB, called NETPACK. NETPACK is a library to model ensembles of neural networks. Ensembles of neural networks tend to yield a much better generalization performance than single neural networks.
NETPACK is freely available for academic usage. Ask us for more information.