Evolving novelty detectors for specific applications

Novelty detection, identifying significant deviations in a systems behaviour, is important in many applications. However, what constitutes novelty is inherently application-specific. Therefore, many existing approaches to novelty detection focus on specific scenarios. Furthermore, approaches shown to generalise over different applications typically require application-specific parameters to be chosen. We propose a system which constructs novelty detectors for specific applications. Neural network-based detectors, with properties taken from dynamic predictive coding, are constructed with methods based on NeuroEvolution of Augmenting Topologies (NEAT). We demonstrate the system over two use-cases, where it outperforms a specialist approach in each case.