Distributed Probabilistic Inference
Large-scale networks of sensing devices are becoming increasingly pervasive, with applications ranging from sensor networks and mobile robot teams to emergency response systems. Often, nodes in these networks need to perform probabilistic inference to combine local, noisy observations into a global, joint estimate of the system state. A simple approach is to collect the data to a central location, where the processing is performed. Yet, collecting all the observations is often impractical in large networks, especially if the nodes communicate over a wireless network and in online control applications, where nodes need estimates of the state in order to make decisions. Suppose instead that each node carries a fragment of the model, and the nodes collaborate to solve the inference task in a distributed manner. In this project, we examine the data structures and algorithms that allow the nodes compute multi-variate solutions in presence of communication delays and node failures.
It is often useful to distributed inference in static systems, where the hidden state does not change over time. For example, in order to effectively control the temperature in a building, the nodes may need to estimate the bias of each temperature sensor at a fixed point in time. We developed a representation that allows the nodes to perform static probabilistic inference with only a local view of the joint probability model. At any point during the execution of the algorithm, the nodes can form a principled approximation of the marginal distribution they wish to compute. The partial results obtained by our algorithm are guaranteed convergence to the correct global solution.
In many cases, we wish to estimate quantities in dynamic settings where we the hidden system state evolves with some known transition model. For example, in a camera network, sensors may wish combine their observations to track the target while estimating their own positions. Because of communication delays and because the nodes can maintain only a limited view of the past, there may be inconsistency among the node beliefs. We propose an effective formulation that lets the nodes perform dynamic inference and recover from any inconsistencies that may arise.
Robust Probabilistic Inference in Distributed Systems
Mark Paskin and Carlos Guestrin, UAI 2004.
A Robust Architecture for Distributed Inference in Sensor Networks
Mark Paskin, Carlos Guestrin, and Jim McFadden, IPSN 2005.
Distributed Inference in Dynamical Systems
Stanislav Funiak, Carlos Guestrin, Mark Paskin, and Rahul Sukthankar, NIPS 2006.