ERA: Learning planner knowledge in complex, continuous and noisy environments

Knowledge acquisition is a critical task in the development of efficient, real-world planning applications. A planner requires knowledge about the objects in the world, its actions and how these actions interact with the objects. Current knowledge acquisition tools for planners are not sufficient to deal with complex, continuous and noisy real-world environments. In particular, in this thesis we focus on the knowledge acquisition needs for mobile robot domains. These domains pose several challenges for learning action models. First, the system must be capable of dealing with continuous variables in the post-conditions and effects of actions. A given action may have multiple effects and each effect may have a range of possible outcomes that depend on the state in which it is. Finally, the robot's noisy sensors will prevent it from detecting the state it is in with certainty. This could result in incorrect data in the training set.; We address these challenges by developing a new approach to learning action models. We call the approach ERA as it consists of Exploration (to gather data) and learning to produce Action models. We evaluate several candidate machine learning techniques, including a new technique developed in this thesis for multi-variate regression tree induction, suitable for supporting knowledge acquisition in mobile robotic domains and develop techniques for incorporating the results of these machine learning techniques as action models within a planning system. We demonstrate the effectiveness of ERA on many different problems from two distinct mobile robot domains including an application on an outdoor mobile robot. In this latter application, our system allows an outdoor robot to learn action models for its navigation actions through experience. To demonstrate the robot's ability to use these action models for planning, a planner has been created that generates high-level navigation plans given the learned action models.