Book Description
Knowledge-based approaches to planning and control offer benefits over classical techniques in applications that involve large yet structured state spaces. However, knowledge bases are time consuming and costly to construct. In this dissertation I introduce a framework for analytical learning that enables the agent to acquire generalizable, domain-specific procedural knowledge in the form of goal-indexed hierarchical task networks by observing a small number of successful demonstrations of goal-driven tasks. I discuss how, in contrast with most algorithms for learning by observation, my approach can learn from unannotated input demonstrations by automatically inferring the purpose of each solution step using the background knowledge about the domain. I discuss the role of hierarchical structure, distributed applicability conditions, and goals in the generalizability of the acquired knowledge. I also introduce an approach for adaptively determining the structure of the acquired knowledge that strikes a balance between generality and operationality, and for making the algorithm robust to changes in the structure of background knowledge. This involves resolving interdependencies among goals using temporal information. I present experimental studies on a number of domains which demonstrate that the quality of acquired knowledge is comparable to handcrafted content in terms of both coverage and complexity. In closing, I review related work and directions for future research.