Multiobjective Optimal Control With Safety as a Priority

This paper develops a lexicographic approach to multiobjective optimal control on models for cyber-physical systems, encompassing in particular stochasticity, limited access to model variables (partial observations), and possibly hybrid (continuous and discrete) dynamics (with the finite-state partially observable Markov decision process framework as a known special instance). The technique is showcased in two new case studies in the area of smart buildings. Technically, the main achievements of this paper are as follows: the application of the lexicographic framework to multiobjective optimization including quantitative probabilistic safety requirements, thus leading to a principled and scalable integration of correct-by-design synthesis for safety and optimal synthesis for performance, the novel extension of the lexicographic framework to partially observed stochastic models with continuous (possibly hybrid) dynamics, and the emphasis on computational aspects, including the use of compact and approximate representations of value functions combined with the quantification of error bounds on model abstractions.