Decentralized adaptation in sensor networks: Analysis and application of regret-based algorithms

We describe a class of decentralized, game theoretic adaptive algorithms which can be deployed to manage sensor activities with low coordination overhead. This class includes traditional game theoretic algorithms such as fictitious play as well as new dynamically adaptive regret matching algorithms, which allow sensors to track a competitively optimal (correlated equilibrium) set of behaviour as it evolves in time. Two applications are given, to a ZigBee-enabled unattended ground sensor network for intruder monitoring, and to a dynamic spectrum allocation scheme for wireless sensor communication.