Performance monitoring of controlled systems under uncertainty

The increasing trend towards delegating complex tasks to artificial agents in safety-critical socio-technical systems makes its performance monitoring of vital importance. In this work a probabilistic approach to online monitoring is proposed on the basis of optimal action selection and Twin Gaussian processes (TGP). A metric built upon the Kullback-Leibler (KL) distance is used to compute the difference between an optimally controlled stochastic process with respect to its specification. The specification is obtained using linearly solvable Markov decision processes (LSMDP). To this end, the Bellman fundamental equation is linearized through an exponential transformation, which allows obtaining the optimal control policy in an explicit manner. Glucose regulation in diabetic patients is used to illustrate the proposed performance monitoring approach.