Interleaved Inductive-Abductive Reasoning for Learning Complex Event Models

We propose an interleaved inductive-abductive model for reasoning about complex spatio-temporal narratives. Typed Inductive Logic Programming (Typed-ILP) is used as a basis for learning the domain theory by generalising from observation data, whereas abductive reasoning is used for noisy data correction by scenario and narrative completion thereby improving the inductive learning to get semantically meaningful event models. We apply the model to an airport domain consisting of video data for 15 turn-arounds from six cameras simultaneously monitoring logistical processes concerned with aircraft arrival, docking, departure etc and a verbs data set with 20 verbs enacted out in around 2500 vignettes. Our evaluation and demonstration focusses on the synergy afforded by the inductive-abductive cycle, whereas our proposed model provides a blueprint for interfacing common-sense reasoning about space, events and dynamic spatio-temporal phenomena with quantitative techniques in activity recognition.