Knowledge-Based Planning – Where does the Knowledge Come From?

Blog post under development. Ideas and thoughts.

There has been a frequent criticism of knowledge-based approaches to planning, including Hierarchical Task Network (HTN) style planning, as to where the encoded processes, descriptions of tasks and agent capabilities, domain constraints, etc come from. I.e. where does the knowledge come from? Its sometimes called the “Knowledge Bottleneck”.

While much of the knowledge encoded in the operators or activity descriptions in early hierarchical task network planners was hand encoded, there has been a lot of work on knowledge capture and domain description methods and tools over the years. Some involved taking authenticated manuals for a domain (such as search and rescue) and extracting the often carefully documented processes involved.

There has been a recent rise in popularity, scope and power of Large Language Models – essentially data extracted from a large corpus (often from internet and web sources) and used to train systems which can generate a range of outputs based on that data. Moe attention is rightly being paid now to the veracity of the results they produce, their biases, their provenance, and their maintenance. There is far to go. But an intriguing possibility emerges… one that needs thought and care ahead of time to ensure its truly useful. And it could be part of the solution to the knowledge bottleneck for knowledge-based planning.

Extract and represent in a shared, reusable form the process knowledge from published manuals, guides, recipes, repair instructions accessible o the internet or world Wide Web.

We have previously used manuals prepared for a community (such as Search and Rescue) as the basis for hand coding the task networks needed for our HTN planners (O-Plan and I-X/I-Plan).

Metadata and Annotations

It is essential that the knowledge represented can be used with permission, reliably and expandability.

Data Sources

There are very many sources of the data that could be obtained. Some initial rules might help guide reliable and maintainable knowledge bases.

  1. Only use sources that are permitted by their licence. The source must have a creditable citation as well as what might be a temporary physical or digital location (library sources, web URL, etc). The original material, the citation, the licence and any current location must be maintained alongside the data itself for any future checking, repeatability , etc.
  2. The date of use, version and any update checks must be maintained in the meta data.
  3. Consider the frequency of update, checks, ways to have community feedback, corrections, etc.
  4. A lot more to come…

Possible Target Representations

While any good representation of the knowledge gained should be easily transformed into future emerging representations and use standards as they emerge, there are some potential initial targets using what is already available on known. Here are some thoughts from our own decades of research on plan representation, shared planning representations which computer systems, people and robots can share, rationale capture in plans, underlying core ontologies for describing activity and agent capability, semantic web representations of processes, etc.

  • OWL/OWL-S – W3C Ontology Web language – Services.
  • SPAR – (DARPA) Shared Planning and Activity Representation.
  • ISO 18629/NIST PSL – Process Specification Language.
  • <I-N-C-A> – Issues, Noes, Constraints and Annotations Ontology.

Domain and Plan Knowledge Capture Work

TBA. Needs reference to our paper on extracting processes from Search and Rescue Manuals.



ISO 18629 –


O-Plan and I-X/I-Plan.

Tate, A. (2003) : a Shared Model for Mixed-initiative Synthesis Tasks, Proceedings of the Workshop on Mixed-Initiative Intelligent Systems (MIIS) at the International Joint Conference on Artificial Intelligence (IJCAI-03), pp. 125-130, Acapulco, Mexico, August 2003.

Tate, A., Dalton. J. and Levine, J. Multi-Perspective Planning – Using Domain Constraints to Support the Coordinated Development of Plans, O-Plan Final Technical Report AFRL-IF-RS-TR-1999-60 April 1999.

Polyak, S. and Tate, A. (1998) Rationale in Planning: Causality, Dependencies and Decisions, The Knowledge Engineering Review, Vol 13(3), September, pp. 247-262, 1998.

Tate, A., Wickler, G., McCluskey, T.L. and Chrpa, L. (2012) Machine Learning and Adaptation of Domain Models to Support Real Time Planning in Autonomous Systems – Month 6 Report, HedLAMP Project Report, University of Edinburgh and University of Huddersfield, 31st August 2012.

Tate, A. (2002) Personal Help Device (PHD) and the Safety Net – a personal agent and its associated local, regional, national and international infrastructure for Aid and Rescue, Papers for the UK Computing Research Committee Workshop on Grand Challenges for Computer Science, National e-Science Centre (NeSC), Edinburgh, Scotland, 25/26 November 2002.

Siebra, Clauirton de Albuquerque and Lino, Natasha Correia Queiroz. Aspects of planning support for human-agent coalitions. J. Braz. Comp. Soc. 2009, vol.15, n.4, pp. 41-55. December 2009.

Tate, A., Buckingham Shum, S.J., Dalton, J, Mancini, C. and Selvin, A.M. (2006) Co-OPR: Design and Evaluation of Collaborative Sensemaking and Planning Tools for Personnel Recovery, Open University Knowledge Media Institute, Technical Report KMI-06-07, March 2006.

This entry was posted in AI and tagged , . Bookmark the permalink.