Learning to Detect Complex Events with Expert Advice

Systems for symbolic event recognition detect occurrences of events in streaming input using a set of event patterns in the form of temporal logical rules. Algorithms for online learning/revising such patterns should be capable of updating the current event pattern set without compromising the quality of the provided service, i.e. the system’s online predictive performance. Towards this, we present an approach based on Prediction with Expert Advice. The experts in our approach are logical rules representing event patterns, which are learnt online via a single-pass strategy. To handle the dynamic nature of the task, an Event Calculus-inspired prediction/event detection scheme allows to incorporate commonsense principles into the learning process.We present a preliminary empirical assessment with promising results.