ResearchSpace

Context-based online policy instantiation for multiple tasks and changing environments

Show simple item record

dc.contributor.author Rosman, Benjamin S
dc.date.accessioned 2015-01-14T05:55:34Z
dc.date.available 2015-01-14T05:55:34Z
dc.date.issued 2014
dc.identifier.citation Rosman, B.S. 2014. Context-based online policy instantiation for multiple tasks and changing environments. In: Pattern Recognition Association of South Africa (PRASA)/RobMech/6th Workshop on African Language Technology (AfLaT), Cape Town, 27-28 November 2014 en_US
dc.identifier.uri http://www.prasa.org/index.php/2012-03-07-10-55-15
dc.identifier.uri http://hdl.handle.net/10204/7835
dc.description Pattern Recognition Association of South Africa (PRASA)/RobMech/6th Workshop on African Language Technology (AfLaT), Cape Town, 27-28 November 2014 en_US
dc.description.abstract This paper addresses the problem of online decision making in continually changing and complex environments, with inherent incompleteness in models of change. A fully general version of this problem is intractable but many interesting domains are rendered manageable by the fact that all instances of a task can be generated from a finite set of qualitatively meaningful contexts. We present an approach to online decision making that exploits this decomposability in a two part procedure. In a task independent exploratory process, our algorithm running on an autonomous agent learns the set of structural landmark contexts which compose its domain, and reduces this set through the use of the symmetry structure of permutation groups. To each reduced landmark we then associate a set of policies independent of global context. This enables an efficient online policy instantiation process that composes from already learnt policy templates. This is illustrated on a spatial navigation domain where the learning agent is shown to be able to play a pursuit-evasion game in random environments with unknown dynamic obstacles. en_US
dc.language.iso en en_US
dc.publisher PRASA en_US
dc.relation.ispartofseries Workflow;13978
dc.subject Online decision making en_US
dc.subject Spatial navigation domains en_US
dc.subject Markov decision process en_US
dc.subject Complex environments en_US
dc.subject Local models en_US
dc.subject Bounded reasoning en_US
dc.subject Autonomous agents en_US
dc.title Context-based online policy instantiation for multiple tasks and changing environments en_US
dc.type Conference Presentation en_US
dc.identifier.apacitation Rosman, B. S. (2014). Context-based online policy instantiation for multiple tasks and changing environments. PRASA. http://hdl.handle.net/10204/7835 en_ZA
dc.identifier.chicagocitation Rosman, Benjamin S. "Context-based online policy instantiation for multiple tasks and changing environments." (2014): http://hdl.handle.net/10204/7835 en_ZA
dc.identifier.vancouvercitation Rosman BS, Context-based online policy instantiation for multiple tasks and changing environments; PRASA; 2014. http://hdl.handle.net/10204/7835 . en_ZA
dc.identifier.ris TY - Conference Presentation AU - Rosman, Benjamin S AB - This paper addresses the problem of online decision making in continually changing and complex environments, with inherent incompleteness in models of change. A fully general version of this problem is intractable but many interesting domains are rendered manageable by the fact that all instances of a task can be generated from a finite set of qualitatively meaningful contexts. We present an approach to online decision making that exploits this decomposability in a two part procedure. In a task independent exploratory process, our algorithm running on an autonomous agent learns the set of structural landmark contexts which compose its domain, and reduces this set through the use of the symmetry structure of permutation groups. To each reduced landmark we then associate a set of policies independent of global context. This enables an efficient online policy instantiation process that composes from already learnt policy templates. This is illustrated on a spatial navigation domain where the learning agent is shown to be able to play a pursuit-evasion game in random environments with unknown dynamic obstacles. DA - 2014 DB - ResearchSpace DP - CSIR KW - Online decision making KW - Spatial navigation domains KW - Markov decision process KW - Complex environments KW - Local models KW - Bounded reasoning KW - Autonomous agents LK - https://researchspace.csir.co.za PY - 2014 T1 - Context-based online policy instantiation for multiple tasks and changing environments TI - Context-based online policy instantiation for multiple tasks and changing environments UR - http://hdl.handle.net/10204/7835 ER - en_ZA


Files in this item

This item appears in the following Collection(s)

Show simple item record