Early detection of abnormal emergent behaviour

Abstract

Emergent behaviour has become a plague of automation systems based on communication networks. Centralized monitoring of the network comes generally to late to suppress unwanted behaviour. It is required to mark the tendency towards state changes in a decentralized manner. The paper discusses the role of local awareness by inspection of the model learning behaviour of feed-forward networks. The correlated movement of weight changes over time provides a clear indication of such profound changes, as demonstrated by some initial experience in industrial automation.