An Explanation for the Efficiency of Scale Invariant Dynamics of Information Fusion in Large Teams Robin Glinton, Paul Scerri, Katia Sycara Robotics Institute Carnegie Mellon University rglinton,pscerri,katia@cs.cmu.edu Abstract – Large heterogeneous teams will often be in sit- uations where sensor data that is uncertain and conflicting is shared across a peer-to-peer network. Not every team member will have direct access to sensors and team mem- bers will be influenced mostly by teammates with whom they communicate directly. Simple models of a large team shar- ing beliefs to reach conclusions about the world show that the dynamics of such belief sharing systems are character- ized by information cascades. These are ripples of belief changes through the system caused by a single additional sensor reading. Glinton et al. [1] showed that such a sys- tem will exhibit qualitatively different dynamics sensitive to ranges over system parameters. In addition they showed that In one particular range, the system exhibits behavior known as scale-invariant dynamics which was found empirically to correspond to dramatically more accurate conclusions be- ing reached by team members. In this paper we provide an analytical explanation for the performance of scale invari- ant dynamics by leveraging signal processing concepts. We show that scale invariant dynamics behave as an adaptive information filter with a response that automatically adjusts to the accuracy of sensor inputs. This adaptation causes the performance gain. Keywords: Self-organization, Complex systems, Large Scale Information Fusion 1 Introduction 1 Large heterogeneous teams will often be in situations where sensor data that is uncertain and conflicting is shared across a peer-to-peer network. Not every team member will have direct access to sensors and team members will be in- fluenced mostly by teammates with whom they communi- cate directly. The effective sharing and use of uncertain information is key to the success of large heterogeneous teams in complex environments because without a correct understanding of the environment it is not possible to appro- priately plan and act. Typically, noisy information is col- lected by some portion of the team and shared via the social 1 This research has been sponsored in part by AFOSR FA95500810356. and/or physical networks connecting members of the team [2]. Each team member will use incoming, uncertain in- formation and the beliefs of those around them to develop their own beliefs about relevant facts. However, the vol- ume of incoming data relative to bandwidth constraints, will often make it impossible for agents to communicate all the received information. Each agent must filter and abstract the information, communicating only its conclusions. Example applications of such systems include large scale disaster re- lief, environmental monitoring, military crisis response etc. [3]. Before such teams are deployed in domains where there are significant costs for bad behavior, it is important to un- derstand and, if necessary, mitigate any system-wide phe- nomena that occur during belief propagation. Understand- ing the dynamics of the system and linking this understand- ing to overall system performance is difficult since network- based belief propagation in large heterogeneous teams ex- hibits complex emergent behaviors [4]. Previous attempts to describe the information dynamics of complex systems in- cludes describing propagation of fads [5, 4], rumors [6] and gossip[7] through social networks. The key difference be- tween this work and previous work is that in previous work a single type of information spread whereas here we can have conflicting data that fundamentally changes the dynamics. Moreover, we are able to use agents to predict and control system dynamics in order to guide the team to areas of opti- mized performance. To analyze the dynamics, we follow Glinton et al. [1] and model a team as being connected via a network with some team members having direct access to sensors and others re- lying solely on neighbors in the network to inform their be- liefs. Each agent uses inference over communications from direct neighbors and sensor data to maintain belief about the environment. This model is attractive because the level of abstraction of the model allows for investigation of team level phenomena decoupled from the noise of high fidelity models or the real-world, allowing for repeatability and sys- tematic varying of parameters. In this model the dynamics of information exchange are dominated by large cascades