Transferring Knowledge from Another Domain for Learning Action Models

Abstract

Learning action models is an important and difficult task for AI planning, since it is both time-consuming and tedious for a human to encode the action models by hand using a formal language such as PDDL. In this paper, we present a new algorithm to learn action models from plan traces by transferring useful knowledge from another domain whose action models are already known. We call this algorithm t-LAMP, (transfer Learning Action Models from Plan traces) which can learn action models in PDDL language with quantifiers from plan traces where the intermediate states can contain noise and partial information. We apply Markov Logic Network to enable knowledge transfer, and show that using the transfer learning framework, the quality of the learned action models are generally better than the case when not using an existing domain for transfer.