This thesis develops a Dynamic Belief Network (DBN) system, based on probabilistic belief networks, for monitoring discrete environments, motivated by the problem of monitoring the movement of robot vehicles and people in a restricted environment, using observation data from light beam sensors. Several alternative network structures are developed which predict the expected positions of objects in the environment; the sensor data is used to update the estimates, outputting beliefs (or probability distributions) for the positions of the moving objects in the environment. The system compares an observed trajectory with a scheduled trajectory, distinguishing between spatial and temporal deviations.
The basic monitoring DBN is provided with a mechanism for handling noisy or incorrect data. The status of a sensor is represented explicitly by an invalidating node providing a symbolic explanation of bad data as being caused by defective sensor. A solution to the Data Association Problem for the domain is developed which uses both externally provided information such as a model of a robot vehicle's motion and its schedule, and internally inferred information about its previous movement. Weight nodes maintain a limited history of the movement of an object and bias the predicted movements according to models of the object movement relating to frequency of movement, inertia (in the direction of movement) and adherence to a schedule.
General inference for such multiply-connected networks is NP-hard. Analysis of the performance of the DBN inference shows that it depends critically on the choice of network structure and the environment specifications. Domain independent meta-level reasoning techniques which use only the network's inferred beliefs, such as pruning and belief threshold assumptions, significantly improve the performance of the inference system.