Avtonomno modeliranje robotskih akcij z odkrivanjem abstraktnih konceptov

Aljaž Košmerlj (2013) Avtonomno modeliranje robotskih akcij z odkrivanjem abstraktnih konceptov. PhD thesis.

Preview

PDF Download (2022Kb)

Abstract

The thesis presents a novel approach to autonomous learning of STRIPS-like robot action models that contain newly discovered abstract concepts. A learnt concept is abstract if it is not explicitly expressed in the learning data (eg. movability, “lower than”, weighted, etc.). We developed STRUDEL, a method that clusters robot actions into groups and then discovers abstract concepts as conditions that classify actions to their respective effects. The definitions of these concepts are used to build the final action model. Both the action model and the concept definitions can be expressed recursively. A new clustering algorithm ACES is used for clustering in STRUDEL. ACES clusters logical action descriptions by similarity of ther structure. Specifically it uses matching of graphs constructed from literals in the action effect descriptions as a distance measure. Performance of STRUDEL and ACES are demonstrated on several experimental domains. Finally, we describe HYPER/CA, a new inductive logic programming (ILP) system, developed as an upgrade of the HYPER system. HYPER/CA was developed to handle learning from noisy data in STRUDEL. We compare it with state-of-the-art ILP systems on several typical ILP learning tasks. Results on the test problems show that HYPER/CA, though quite slower then the algorithms used in the comparison, can attain similar accuracy while building smaller hypotheses.

Actions (login required)