This project focuses on how a supervised classification method can be implemented in a decision-making framework and how this can provide knowledge about the expected utility of the classification, with this also misclassification of Type I and Type II errors.

The problem is exemplified through an experimental platform of a simple pendulum. The pendulum is a physical system that exhibits a dynamic behavior over time that depends on several characteristics. A pendulum might appear to be a somewhat artificial system, but depending on the chosen system parameters may represent many phenomena of socio-ecological systems, such as climate variations, variations in traffic, or responses of structural and mechanical systems.

A probabilistic model is developed for the representation of the different information states. The already developed database of realizations of system parameters and system behaviors is augmented to reflect observations of the system behaviors, including the effect of information states based on the probabilistic model.

The experimental platform establishes a foundation for a machine-learning classification of the observations into different information states. The information states will be predicted with adequate precision to reflect the decision-makers preferences regarding Type I and Type II errors. The choices involved in the above-outlined approach can be identified by utilization of the concept of Value of Information. The Value of Information may now be applied to obtain the Value of Classification, based on the expected value gain of utilizing a classification scheme.