Planning in BDI Agent Systems

Planning in BDI Agent Systems. de Silva, L. Ph.D. Thesis, School of Computer Science and Information Technology, RMIT University, 2009.

Dissertation abstract bibtex

Belief-Desire-Intention (BDI) agent systems are a popular approach to developing agents for complex and dynamic environments. These agents rely on context sensitive expansion of plans, acting as they go, and consequently, they do not incorporate a generic mechanism to do any kind of “look-ahead” or offline planning. This is useful when, for instance, important resources may be consumed by executing steps that are not necessary for a goal; steps are not reversible and may lead to situations in which a goal cannot be solved; and side effects of steps are undesirable if they are not useful for a goal. In this thesis, we incorporate planning techniques into BDI systems.

First, we provide a general mechanism for performing “look-ahead” planning, using Hierarchical Task Network (HTN) planning techniques, so that an agent may guide its selection of plans for the purpose of avoiding negative interactions between them. Unlike past work on adding such planning into BDI agents, which do so only at the implementation level without any precise semantics, we provide a solid theoretical basis for such planning.

Second, we incorporate first principles planning into BDI systems, so that new plans may be created for achieving goals. Unlike past work, which focuses on creating low-level plans, losing much of the domain knowledge encoded in BDI agents, we introduce a novel technique where plans are created by respecting and reusing the procedural domain knowledge encoded in such agents; our abstract plans can be executed in the standard BDI engine using this knowledge. Furthermore, we recognise an intrinsic tension between striving for abstract plans and, at the same time, ensuring that unnecessary actions, unrelated to the specific goal to be achieved, are avoided. To explore this tension, we characterise the set of “ideal” abstract plans that are non-redundant while maximally abstract, and then develop a more limited but feasible account where an abstract plan is “specialised” into a plan that is non-redundant and as abstract as possible. We present theoretical properties of the planning frameworks, as well as insights into their practical utility.

@PhdThesis{deSilva2009PhDThesis,
  author    = {Lavindra de Silva},
  title     = {Planning in BDI Agent Systems},
  school    = {School of Computer Science and Information Technology, RMIT University},
  year      = {2009},
  abstract  = {Belief-Desire-Intention (BDI) agent systems are a popular approach to developing agents for complex and dynamic environments. These agents rely on context sensitive expansion of plans, acting as they go, and consequently, they do not incorporate a generic mechanism to do any kind of “look-ahead” or offline planning. This is useful when, for instance, important resources may be consumed by executing steps that are not necessary for a goal; steps are not reversible and may lead to situations in which a goal cannot be solved; and side effects of steps are undesirable if they are not useful for a goal. In this thesis, we incorporate planning techniques into BDI systems.<br/><br/>

  First, we provide a general mechanism for performing “look-ahead” planning, using Hierarchical Task Network (HTN) planning techniques, so that an agent may guide its selection of plans for the purpose of avoiding negative interactions between them. Unlike past work on adding such planning into BDI agents, which do so only at the implementation level without any precise semantics, we provide a solid theoretical basis for such planning.<br/><br/>

  Second, we incorporate first principles planning into BDI systems, so that new plans may be created for achieving goals. Unlike past work, which focuses on creating low-level plans, losing much of the domain knowledge encoded in BDI agents, we introduce a novel technique where plans are created by respecting and reusing the procedural domain knowledge encoded in such agents; our abstract plans can be executed in the standard BDI engine using this knowledge. Furthermore, we recognise an intrinsic tension between striving for abstract plans and, at the same time, ensuring that unnecessary actions, unrelated to the specific goal to be achieved, are avoided. To explore this tension, we characterise the set of “ideal” abstract plans that are non-redundant while maximally abstract, and then develop a more limited but feasible account where an abstract plan is “specialised” into a plan that is non-redundant and as abstract as possible. We present theoretical properties of the planning frameworks, as well as insights into their practical utility.},
  url_Dissertation = {https://researchrepository.rmit.edu.au/esploro/outputs/doctoral/Planning-in-BDI-agent-systems/9921861322501341?institution=61RMIT_INST}
}

Downloads: 0

{"_id":"M7vBnZM2rsEZpxxjn","bibbaseid":"desilva-planninginbdiagentsystems-2009","authorIDs":[],"author_short":["de Silva, L."],"bibdata":{"bibtype":"phdthesis","type":"phdthesis","author":[{"firstnames":["Lavindra"],"propositions":["de"],"lastnames":["Silva"],"suffixes":[]}],"title":"Planning in BDI Agent Systems","school":"School of Computer Science and Information Technology, RMIT University","year":"2009","abstract":"Belief-Desire-Intention (BDI) agent systems are a popular approach to developing agents for complex and dynamic environments. These agents rely on context sensitive expansion of plans, acting as they go, and consequently, they do not incorporate a generic mechanism to do any kind of “look-ahead” or offline planning. This is useful when, for instance, important resources may be consumed by executing steps that are not necessary for a goal; steps are not reversible and may lead to situations in which a goal cannot be solved; and side effects of steps are undesirable if they are not useful for a goal. In this thesis, we incorporate planning techniques into BDI systems. First, we provide a general mechanism for performing “look-ahead” planning, using Hierarchical Task Network (HTN) planning techniques, so that an agent may guide its selection of plans for the purpose of avoiding negative interactions between them. Unlike past work on adding such planning into BDI agents, which do so only at the implementation level without any precise semantics, we provide a solid theoretical basis for such planning. Second, we incorporate first principles planning into BDI systems, so that new plans may be created for achieving goals. Unlike past work, which focuses on creating low-level plans, losing much of the domain knowledge encoded in BDI agents, we introduce a novel technique where plans are created by respecting and reusing the procedural domain knowledge encoded in such agents; our abstract plans can be executed in the standard BDI engine using this knowledge. Furthermore, we recognise an intrinsic tension between striving for abstract plans and, at the same time, ensuring that unnecessary actions, unrelated to the specific goal to be achieved, are avoided. To explore this tension, we characterise the set of “ideal” abstract plans that are non-redundant while maximally abstract, and then develop a more limited but feasible account where an abstract plan is “specialised” into a plan that is non-redundant and as abstract as possible. We present theoretical properties of the planning frameworks, as well as insights into their practical utility.","url_dissertation":"https://researchrepository.rmit.edu.au/esploro/outputs/doctoral/Planning-in-BDI-agent-systems/9921861322501341?institution=61RMIT_INST","bibtex":"@PhdThesis{deSilva2009PhDThesis,\n author = {Lavindra de Silva},\n title = {Planning in BDI Agent Systems},\n school = {School of Computer Science and Information Technology, RMIT University},\n year = {2009},\n abstract = {Belief-Desire-Intention (BDI) agent systems are a popular approach to developing agents for complex and dynamic environments. These agents rely on context sensitive expansion of plans, acting as they go, and consequently, they do not incorporate a generic mechanism to do any kind of “look-ahead” or offline planning. This is useful when, for instance, important resources may be consumed by executing steps that are not necessary for a goal; steps are not reversible and may lead to situations in which a goal cannot be solved; and side effects of steps are undesirable if they are not useful for a goal. In this thesis, we incorporate planning techniques into BDI systems. \n\n First, we provide a general mechanism for performing “look-ahead” planning, using Hierarchical Task Network (HTN) planning techniques, so that an agent may guide its selection of plans for the purpose of avoiding negative interactions between them. Unlike past work on adding such planning into BDI agents, which do so only at the implementation level without any precise semantics, we provide a solid theoretical basis for such planning. \n\n Second, we incorporate first principles planning into BDI systems, so that new plans may be created for achieving goals. Unlike past work, which focuses on creating low-level plans, losing much of the domain knowledge encoded in BDI agents, we introduce a novel technique where plans are created by respecting and reusing the procedural domain knowledge encoded in such agents; our abstract plans can be executed in the standard BDI engine using this knowledge. Furthermore, we recognise an intrinsic tension between striving for abstract plans and, at the same time, ensuring that unnecessary actions, unrelated to the specific goal to be achieved, are avoided. To explore this tension, we characterise the set of “ideal” abstract plans that are non-redundant while maximally abstract, and then develop a more limited but feasible account where an abstract plan is “specialised” into a plan that is non-redundant and as abstract as possible. We present theoretical properties of the planning frameworks, as well as insights into their practical utility.},\n url_Dissertation = {https://researchrepository.rmit.edu.au/esploro/outputs/doctoral/Planning-in-BDI-agent-systems/9921861322501341?institution=61RMIT_INST}\n}\n\n\n","author_short":["de Silva, L."],"key":"deSilva2009PhDThesis","id":"deSilva2009PhDThesis","bibbaseid":"desilva-planninginbdiagentsystems-2009","role":"author","urls":{" dissertation":"https://researchrepository.rmit.edu.au/esploro/outputs/doctoral/Planning-in-BDI-agent-systems/9921861322501341?institution=61RMIT_INST"},"metadata":{"authorlinks":{}}},"bibtype":"phdthesis","biburl":"hierarchical-task.net/bibtex/theses-in-hierarchical-planning.bib","creationDate":"2020-04-26T03:45:17.871Z","downloads":1,"keywords":[],"search_terms":["planning","bdi","agent","systems","de silva"],"title":"Planning in BDI Agent Systems","year":2009,"dataSources":["vY5emPje84NjvCwk9"]}