Inductive Generalization of Analytically Learned Goal Hierarchies

We describe a new approach for learning procedural knowledge represented as Teleoreactive Logic Programs (TLPs) using relational behavior traces as input. A TLP organizes task decomposition skills hierarchically and attaches them explicitly defined goals. Our approach integrates analytical learning with inductive generalization in order to learn these skills. The analytical component predicts the goal dependencies given a successful behavior trace and produces the structure of the skill hierarchy. The inductive component generates successful and failed traces using those skills and refines their applicability conditions with an ILP component. We claim that explicit goal representation of TLPs allows the learning system to determine positive and negative examples used in learning. We claim that our approach uses less expert effort compared to a purely inductive approach and may perform better compared to a purely analytical approach.