Probabilistic Backward and Forward Reasoning in Stochastic Relational Worlds

Inference in graphical models has emerged as a promising technique for planning. A recent approach to decision-theoretic planning in relational domains uses forward inference in dynamic Bayesian networks compiled from learned probabilistic relational rules. Inspired by work in non-relational domains with small state spaces, we derive a back-propagation method for such nets in relational domains starting from a goal state mixture distribution. We combine this with forward reasoning in a bidirectional two-filter approach. We perform experiments in a complex 3D simulated desktop environment with an articulated manipulator and realistic physics. Empirical results show that bidirectional probabilistic reasoning can lead to more efficient and accurate planning in comparison to pure forward reasoning.