Jilani, Rabia (2017) Learning Static Knowledge for AI Planning Domain Models via Plan Traces. Doctoral thesis, University of Huddersfield.
Abstract

Learning is fundamental to autonomous behaviour and from the point of view of Machine Learning, it is the ability of computers to learn without being programmed explicitly. Attaining such capability for learning domain models for Automated Planning (AP) engines is what triggered research into developing automated domain-learning systems. These systems can learn from training data. Until recent research it was believed that working in dynamically changing and unpredictable environments, it was not possible to construct action models a priori. After the research in the last decade, many systems have proved effective in engineering domain models by learning from plan traces. However, these systems require additional planner oriented information such as a partial domain model, initial, goal and/or intermediate states. Hence, a question arises - whether or not we can learn a dynamic domain model, which covers all domain behaviours from real-time action sequence traces only.

The research in this thesis extends an area of the most promising line of work that is connected to work presented in an REF Journal paper. This research aims to enhance the LOCM system and to extend the method of Learning Domain Models for AI Planning Engines via Plan Traces. This method was first published in ICAPS 2009 by Cresswell, McCluskey, and West (Cresswell, 2009). LOCM is unique in that it requires no prior knowledge of the target domain; however, it can produce a dynamic part of a domain model from training. Its main drawback is that it does not produce static knowledge of the domain, and its model lacks certain expressive features. A key aspect of research presented in this thesis is to enhance the technique with the capacity to generate static knowledge. A test and focus for this PhD is to make LOCM able to learn static relationships in a fully automatic way in addition to the dynamic relationships, which LOCM can already learn, using plan traces as input.

We present a novel system - The ASCoL (Automatic Static Constraints Learner) which provides a graphical interface for visual representation and exploits directed graph discovery and analysis technique. It has been designed to discover domain-specific static relations/constraints automatically in order to enhance planning domain models. The ASCoL method has wider applications. Combined with LOCM, ASCoL can be a useful tool to produce benchmark domains for automated planning engines. It is also useful as a debugging tool for improving existing domain models. We have evaluated ASCoL on fifteen different IPC domains and on different types of goal-oriented and random-walk plans as input training data and it has been shown to be effective.

Information
Library
Documents
[img]
Preview
FINAL THESIS - Jilani.pdf - Accepted Version
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (3MB) | Preview
Statistics

Downloads

Downloads per month over past year

Add to AnyAdd to TwitterAdd to FacebookAdd to LinkedinAdd to PinterestAdd to Email