Abstract:

The dynamic execution of plans in uncertain domains requires the ability to infer likely current and future world states from past observations. This task can be cast as inference on Dynamic Belief Networks (DBNs) but the resulting networks are difficult to solve with exact methods. We investigate and extend simulation algorithms for approximate inference on Bayesian networks and a propose a new algorithm, called Rewind/Replay, for generating a set of simulations weighted by their likelihood given past observations. We validate our algorithm on a DBN containing thousands of variables, which models the spread of wildfire.