Peter A. Jarvis

Identifying terrorist activity with AI plan recognition technology

We describe the application of plan recognition techniques to support human intelligence analysts in processing national security alert sets by automatically identifying the hostile intent behind them. Identifying the intent enables us to both prioritize and explain the alert sets for succinct user presentation. Our empirical evaluation demonstrates that the approach can handle alert sets of as many as 20 elements and can readily distinguish between false and true alarms. We discuss the important opportunities, for future work, that will increase the cardinality of the alert sets supported by the system to the level demanded by a deployable application. In particular, we outline opportunities to bring the analysts into the process and the opportunities for heuristic improvements to the plan recognition algorithm.

Efficient flexible planning via dynamic flexible constraint satisfaction

Abstract Recent advances in AI planning have centred upon the reduction of planning to a constraint satisfaction problem (CSP) enabling the application of the efficient search algorithms available in this area. This paper continues this approach, presenting a novel technique which exploits (restriction/relaxation-based) dynamic CSP (rrDCSP) in order to further improve planner performance. Using the standard Graphplan framework, it is shown how significant efficiency gains may be obtained by viewing plan extraction as the solution of a hierarchy of such rrDCSPs. Furthermore, by using flexible constraints as a formal foundation, it is shown how the traditional boolean notion of planning can be extended to incorporate prioritised and preference-based information. Plan extraction in this context is shown to generalise the Boolean rrDCSP approach, being systematically supported by the recently developed solution techniques for dynamic flexible CSPs (DFCSPs). The proposed techniques are evaluated via benchmark boolean problems and a novel flexible benchmark problem. Results obtained are very encouraging.

Pointing to places in a deductive geospatial theory

Issues in the description of places are discussed in the context of a logical geospatial theory. This theory lies at the core of the system GeoLogica, which deduces answers to geographical questions based on knowledge provided by multiple agents.

Program Synthesis for Multi-agent Question Answering

Techniques that were developed for program synthesis are being applied to allow multiple agents to communicate with each other and cooperate to solve a single problem. We illustrate the use of program synthesis techniques in a system that answers questions about geography.