Domain Model Acquisition with Missing Information and Noisy Data
Peter J. Gregory, Alan Lindsay, Julie Porteous · 2017
In this work, we address the problem of learning planning domain models from example action traces that contain missing and noisy data. In many situations, the action traces that form the input to domain model acquisition systems are sourced from observations or even natural language descriptions of plans. It is often the case that these observations are noisy and incomplete. Therefore, making domain model acquisition systems that are robust to such data is crucial. Previous approaches to this problem have relied upon having access to the underlying state in the input plans. We lift this assumption and provide a system that does not require any state information. We build upon the LOCM family of algorithms, which also lift this assumption in the deterministic version of the domain model acquisition problem, to provide a domain model acquisition system that learns domain models from noisy plans with missing information.