Decision-theoretic monitoring of cyber-physical systems

Abstract

Runtime monitoring has been proposed as an alternative to formal verification for safety critical systems. This paper introduces a decision-theoretic view of runtime monitoring. We formulate the monitoring problem as a Partially Observable Markov Decision Process (POMDP). Furthermore, we adopt a Partially Observable Monte-Carlo Planning (POMCP) to compute an approximate optimal policy of the monitoring POMDP. We show how to construct the POMCP for the monitoring problem and demonstrate experimentally that it can be effectively applied even when some of the state-space variables are continuous, the case where many other POMDP solvers fail. Experimental results on a mobile robot system show the effectiveness of the proposed POMDP-monitor.