Prediction Process in Multi-Agent System Online Monitoring: Centralized and Distributed Approaches

This paper discusses the prediction process, which is the main step of the online monitoring process for a multi-agent plan. The monitoring process uses a relational model to estimate the internal status of the system, which is dynamic (changes over time). Unfortunately, the agents have partial observability of the environment; thus, the monitoring process cannot accurately determine the system status (known in the literature as belief state) at any instant. The prediction process is composed of two stages: a simulation stage (prediction of all possible system states at the succeeding time) and a clipping stage (elimination of states that are incompatible with the observations or with the constraints from predicted system states).