An architecture for planning in uncertain domains

This paper presents an architecture for intelligent planning in uncertain real domains. This architecture is based on the paradigm of beliefs, desires and intention (BDI). The planning technique consists in the use of Markov Decision Processes (MDP) and Partially Observed MDP (POMDP). The output of the planning consists in advices that the system provides to the operator of a power plant. Specifically, the process experimented in this work is the uncertainty that a classical controller (PID) observes in the control of the level of water of the drum in a steam generator of a power plant. The graphics of the optimal trajectory of control is discretized in order to form a finite set of states in the space. The actions are the increments or decrements of the variables that produce the movements in the control space. The transition matrix is obtained with real data from different operation conditions of the plant. The reward and observation functions are obtained from experimented operators of power plants