Dynamic action spaces for information gain maximization in search and exploration

The problem we investigate is how an autonomous, mobile agent can search for a hidden, moving target efficiently. A good control strategy plans more informative sensing of the world, allowing the agent to find the target quickly. Searching for moving targets typically involves planning over probability distributions, or beliefs, that characterize the possible locations of the target. Most motion strategies choose actions that reduce the uncertainty of the current belief by maximizing the predicted information gain of the next action, but computing good multi-step plans is usually computationally intractable (Sondik, 1971) due to the high dimensionality of the action and belief spaces. In this paper, we describe a novel algorithm for generating search plans using dynamic action spaces. The algorithm clusters a particle filter description of the current belief at each point in time, and uses search to compute a trajectory through the clusters in order to maximize information gain. This model allows us to efficiently compute finite-horizon multi-step plans in extremely high dimensional problems. We show preliminary results for an unknown target tracking problem