A Changepoint Method for Open-World Novelty Detection
Novelty detection in open worlds is a valuable endeavor
for improving the robustness of autonomous systems. Open
worlds are characterized by a lack of constraints on the
types of novelties that might occur. Here we describe recent
progress on detecting novelties in open-world numerical
properties in our OpenMIND agent, a domain-independent
planning-based AI architecture.
Our approach has three elements:
(1) Feature construction by crossing a tractable number
of domain-dependent sensors with some basic domainindependent
statistical derivations, (2) online, univariate
changepoint detection on each signal, and (3) negative feature
selection by screening for false positives in non-novel
scenarios. We report on the effectiveness of this approach
in a single-blind evaluation in which OpenMIND plays a
game and encounters novelties never observed by it or its
developers.