Risk, reliability and machine learning

Contact persons: Arne Bang Huseby, Kristian Ranestad
Keywords: Multistate systems, algebraic methods, monomial ideals, matroids, decision trees, optimization, reinforcement learning
Research group: Risk and Stochastics, Algebra and algebraic geometry
Department of Mathematics

System reliability calculation often requires the integration of various disciplines like engineering, mathematics, computer science, and statistics. This multidisciplinary approach poses a computational challenge, particularly in harmonising different methodologies, models, and data formats into a cohesive reliability assessment framework. The structure of a multicomponent system can be represented in the form of a decision tree, a representation which is well suited for efficient calculations. However, the computational complexity grows by the size of the tree and soon becomes unmanageable. Surrogate tree models obtained using machine learning techniques allow the structure to be approximated by simpler trees, and thus reduce the computational complexity. Recently, the study of multistate systems has gained focus, and innovative approaches involving algebraic methods and monomial ideals have been proposed. Algebraic methods also allow the structure to be decomposed into simpler binary structures.

Within this theme we are interested in project proposal related to state-of-the-art reliability calculation methods. We also welcome proposals within the general field of risk analysis emphasising computational challenges, especially related to optimisation and decision support under uncertainty.

Methodological topics:

  • Algebraic methods and multistate systems
  • Sequential decision problems and reinforcement learning
  • Quality control of autonomous systems
  • Surrogate system models, decision trees and machine learning

Applied topics:

  • Optimizing Energy Efficiency in Machine Learning: Balancing Model Size, Accuracy, and Risk. Industrial use of AI is often based on machine learning models that are expensive (both in terms of cost and energy consumption) to create, deploy and operate. One is therefore interested in ways such models can be simplified without increasing uncertainty beyond what is acceptable. To balance the tradeoff between model size and accuracy, a risk-based approach should be used. This requires quantification of model uncertainties, and the connection of those uncertainties with the consequences of downstream tasks to establish a risk metric. A model may then be optimised with respect to models size while still staying within the acceptable criteria for the risk metric. This will require both developing a suitable risk assessment framework and exploring model simplification techniques.
  • Uncertainty of combined physics-driven and data-driven models for hydrogen safety. With increasing focus on hydrogen as an energy source/carrier, there is a growing need for analysis of system safety, including structural reliability, multicomponent systems, safety barriers, quantitative risk analysis (QRAs), etc. As working with detailed physical models is very time-consuming, there is a need for surrogate models or emulators which can adequately represent the underlying structure, and at the same time speed-up calculations. In order to design such emulators efficiently, machine learning techniques are essential. When creating surrogate models from detailed physics-based possibly combined with data-driven insights from sensor data, the ultimate challenge is to ensure that the prediction uncertainty represents acceptable safety risk. Approaches like uncertainty quantification, machine learning on sensor data from the assets, physical models and risk modelling needs to be addressed within a consistent framework.
  • Machine-readable interpretation of COLREG. The area of autonomous and remotely controlled functions on ships is developing fast with several industrial projects now approaching commercial maturity and some systems already in operation. The rules of the road for the sea, the International Regulation for Preventing Collision at Sea (COLREG) are written by humans for humans and contain several qualitative terms to describe acceptable collision avoidance behavior, such as “ample time”, “readily apparent” and “good seamanship”. A direct translation into machine readable code is therefore not straightforward. A machine-readable interpretation of COLREG, able to provide clear and quantified requirements for given navigational situations, will therefore be of great value to a range of industry actors. To enable industry-wide impact, such an interpretation needs to be accepted by all relevant industry stakeholders, and therefore needs to be unbiased, based on solid sources and a well-documented approach for analysis.

External partners:

  • DNV AS
  • Avinor Group AS
  • Kongsberg Maritime AS
  • Norwegian Computer Center (NR)
  • SINTEF