Using mathematics to make decisions and predictions is central to Quintessa’s activities, and much of our work therefore falls within the scope of AI and machine learning. An example use of AI by Quintessa is in agent-based modelling, where individual agents have been used to simulate human decision making using AI design patterns; the combination of these decisions and their consequences generates the emergent system behaviour, which can be understood and modified both at the individual and system level.
Machine learning can be categorised as either supervised or unsupervised. Supervised learning is used where data have known target attributes to train a method; once trained, the method can then be applied to make predictions for other data where the target attributes are unknown, either in the form of a regression for continuous attributes or a classification for discrete attributes. Quintessa uses these approaches in tandem with other forms of statistical modelling to support its consultancy work, including predictions based on Advanced Gas-cooled Reactor data.
Unsupervised learning identifies patterns in data where the target attributes are unknown. An example of this is clustering analysis, where groups of similar data are identified based on similarities in their attributes. Quintessa has used this approach to help rationalise a large dataset of radionuclide information.
Most machine learning algorithms require large quantities of data; where these are not available, other statistical modelling methods may be more appropriate. Quintessa recognises the strengths and weaknesses of different approaches and we are able to select the most suitable according to the problem. As with all forms of data modelling, the aim is to gain insights into the systems represented by the data, and thus develop robust understanding and predictions. It is therefore important to take a rounded approach to avoid ‘black-box’ and ‘over-fitted’ results.