Action Model Acquisition Using Sequential Pattern Mining

Abstract

This paper presents an approach to learn the agents' action model (action blueprints orchestrating transitions of the system state) from plan execution sequences. It does so by representing intra-action and inter-action dependencies in the form of a maximum satisfiability problem (MAX-SAT), and solving it with a MAX-SAT solver to reconstruct the underlying action model. Unlike previous MAX-SAT driven approaches, our chosen dependencies exploit the relationship between consecutive actions, rendering more accurately learnt models in the end.