This research program will enable radically new capabilities for the US armed forces to deploy intelligent decentralized knowledge learning and planning algorithms for teams of heterogeneous autonomous static and mobile agents. Our research plan is based on the key insight that nonparametric Bayesian models provide a powerful framework for reasoning about objects and relations in settings in which these objects and relations are not predefined. This feature is particularly attractive for missions such as long term persistent surveillance for which it is virtually impossible to specify the size of the model and the number of variables a priori. The research consists of three main thrusts: decentralized inference and model learning; decentralized planning under uncertainty; and information sharing and consensus.

Decentralized Inference and Model Learning

To overcome the fundamental limitations in the expressiveness of existing models, we propose to investigate using nonparametric Bayesian models to handle noisy data, bias, occlusion and heterogeneous data. Our primary research focus in this thrust will be to extend the nonparametric Bayesian models to decentralized settings, with a particular focus on developing the Bayesian framework for modeling how decentralized agents observe and interrogate their environment, employing localized communication.

Decentralized Planning Under Uncertainty

Our  research will develop new decentralized planning algorithms that directly integrate nonparametric Bayesian models. This research will include robustifying planners to model uncertainty, exploiting the structure of graphical models in stochastic planners, and developing techniques that accurately predict the value of additional measurements in reducing the uncertainty.

Information Sharing and Consensus

Our  research will develop consensus algorithms that enable information sharing for inference, learning, and planning across distributed agents. This work will ensure that these algorithms account for uncertainty and temporal evolution in the local models, and we will leverage the fact that nonparametric Bayesian methods address uncertainty in structural aspects of the model while incorporating the inherent constraints of distributed systems.