Efficient Computation of Sensor Activation Decisions in Discrete-Event Systems

This paper considers partially-observed discrete-event systems where sensors are associated with events observable to an agent monitoring the system. The agent is capable of turning the sensors for events on and off dynamically, depending on the trajectory of the system. Reading data from the sensors may be costly so it is imperative that their use be reduced for reasons such as energy, bandwidth or security. When a sensor for an event is on / active any occurrence of the event is detected by the agent and is not detected otherwise. The agent may employ different sensor activation policies, depending on the task at hand. Sensor activation policies are defined over the transitions of a state-transition representation of the system. From sensor activation policies a map from observed event sequences to sensor activation decisions can be computed which the agent can use to determine which sensors to turn on / off and when. In this paper, we consider two subclasses of sensor activation policies of increasing generality. For each subclass, we demonstrate ways to compute maps from observed event sequences to sensor activation decisions in polynomial time.