Reinforcement Learning and Graph Neural Networks for Probabilistic Risk Assessment
Joachim Grimstad, Andrey Morozov
TL;DR
The paper addresses the scalability challenge of Probabilistic Risk Assessment (PRA) models by proposing a framework that unifies fault-tree PRA with modern graph-based machine learning. It leverages Graph Neural Networks (GNNs) to capture dependencies and graph structure and Reinforcement Learning (RL), including Proximal Policy Optimization (PPO), to learn node-level quantitative attributes and structural edits on fault-tree graphs. The approach outlines vertex-, edge-, and graph-level methods to estimate failure probabilities, reveal hidden dependencies via link prediction, and modify graph structure to identify Minimal Cut Sets (MCS), potentially enabling generic PRA solvers trained on data or simulations. If successful, this data-driven paradigm could improve scalability, adaptability to dynamic systems, and insight into failure pathways for complex engineering systems.
Abstract
This paper presents a new approach to the solution of Probabilistic Risk Assessment (PRA) models using the combination of Reinforcement Learning (RL) and Graph Neural Networks (GNNs). The paper introduces and demonstrates the concept using one of the most popular PRA models - Fault Trees. This paper's original idea is to apply RL algorithms to solve a PRA model represented with a graph model. Given enough training data, or through RL, such an approach helps train generic PRA solvers that can optimize and partially substitute classical PRA solvers that are based on existing formal methods. Such an approach helps to solve the problem of the fast-growing complexity of PRA models of modern technical systems.