One-shot learning: From domain knowledge to action models

Anonymous

Keywords: model acquisition, knowledge-based learning

TL;DR: Hybrid approach to model acquisition that compensates a lack of available data with domain specific knowledge provided by experts

Abstract: Most approaches to learning action planning models heavily rely on a significantly large volume of training samples or plan observations. In this paper, we adopt a different approach based on deductive learning from domain-specific knowledge, specifically from logic formulae that specify constraints about the possible states of a given domain. The minimal input observability required by our approach is a single example composed of a full initial state and a partial goal state. We will show that exploiting specific domain knowledge enables to constrain the space of possible action models as well as to complete partial observations, both of which turn out helpful to learn good-quality action models.