Framing Situation Prediction as a Sequence Prediction Problem: A Situation Evolution Model Based on Continuous-Time Markov Chains

Following the acknowledged JDL data fusion model, a situation can be characterized as a set of objects in relations. Considering that this object-relational composition may change over time, as the monitored objects may alter their states (such as changing their event type and position), we can summarize a situation's evolution in a high-level fashion by the sequence of object-relational states it has evolved through, i.e., the sequence of its situation states. Thus, we propose to discretize the continuous evolution of the monitored real-world objects into a sequence of their different joint relational states (e.g., defined by the topological or qualitative distance relations holding between those objects), representing our alphabet, and consequently treat the problem of predicting a monitored situation's further qualitative evolution as a sequence prediction problem. We examine this approach on real-world data from the domain of road traffic incident management, in which situations are characterized by a changing aggregate of different event types, denoting the distinct phases of the monitored road traffic incidents. For the problem of predicting the monitored situation's next discrete state, we propose a predictive situation evolution model based on a first-order Continuous-Time Markov Chain. Our proposed structured situation prediction approach is applicable to all problem domains in which situations can be formulated as sequences of discretized states.