Decision support to police violent crime

Explaining the science

The models in this project are based on established Bayesian decision support technologies - which have been successfully applied to a variety of analogous dynamic settings. The three-level model produced enables synthesis of the expert judgments of the domain experts with the streaming, but patchy, real time electronic observations concerning specific suspects.

At the deepest hidden level of this hierarchy lies a new customised family called the reduced dynamic chain event graph (RDCEG) recently developed by the PI for specifically modelling such population processes. These express as graphs an expert’s natural language hypotheses about the various different ways people in certain radicalised states might develop; translating such hypotheses into families of probability models. 

At an intermediate level are further expert judgments about the tasks someone at a particular stage of preparing a violent attack might need to perform. The final surface layer – usually the only one directly, albeit partially observable – describes what people might actually do to perform such tasks. This provides the final probabilistic layer, linking data streams to tasks and hence to states. 

These three levels then synthesise into a single probability model of the whole process. This enables the user to calculate expected utility scores that can then help to guide the appropriate deployment of resources, to ideally deter or diffuse potential attacks, or otherwise to at least frustrate these.