ResearchSpace

Action priors for learning domain invariances

Show simple item record

dc.contributor.author Rosman, Benjamin S
dc.contributor.author Ramamoorthy, S
dc.date.accessioned 2015-08-31T06:51:56Z
dc.date.available 2015-08-31T06:51:56Z
dc.date.issued 2015-04
dc.identifier.citation Rosman, B.S. and Ramamoorthy, S. 2015. Action priors for learning domain invariances. IEEE Transactions of Autonomous Mental Development, vol. 7(2), pp 107-118 en_US
dc.identifier.issn 1943-0604
dc.identifier.uri http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=7079487
dc.identifier.uri http://hdl.handle.net/10204/8114
dc.description Copyright; 2015 IEEE Xplore. Due to copyright restrictions, the attached PDF file only contains the abstract of the full text item. For access to the full text item, please consult the publisher's website. The definitive version of the work is published in IEEE Transactions of Autonomous Mental Development, vol. 7(2), pp 107-118 en_US
dc.description.abstract An agent tasked with solving a number of different decision making problems in similar environments has an opportunity to learn over a longer timescale than each individual task. Through examining solutions to different tasks, it can uncover behavioural invariances in the domain, by identifying actions to be prioritised in local contexts, invariant to task details. This information has the effect of greatly increasing the speed of solving new problems. We formalise this notion as action priors, defined as distributions over the action space, conditioned on environment state, and show how these can be learnt from a set of value functions. We apply action priors in the setting of reinforcement learning, to bias action selection during exploration. Aggressive use of action priors performs context based pruning of the available actions, thus reducing the complexity of look ahead during search. We additionally define action priors over observation features, rather than states, which provides further flexibility and generalisability, with the additional benefit of enabling feature selection. Action priors are demonstrated in experiments in a simulated factory environment and a large random graph domain, and show significant speed ups in learning new tasks. Furthermore, we argue that this mechanism is cognitively plausible, and is compatible with findings from cognitive psychology. en_US
dc.language.iso en en_US
dc.publisher IEEE Xplore en_US
dc.relation.ispartofseries Workflow;15282
dc.subject Search pruning en_US
dc.subject Action selection en_US
dc.subject Action ordering en_US
dc.subject Transfer learning en_US
dc.subject Reinforcement learning en_US
dc.title Action priors for learning domain invariances en_US
dc.type Article en_US
dc.identifier.apacitation Rosman, B. S., & Ramamoorthy, S. (2015). Action priors for learning domain invariances. http://hdl.handle.net/10204/8114 en_ZA
dc.identifier.chicagocitation Rosman, Benjamin S, and S Ramamoorthy "Action priors for learning domain invariances." (2015) http://hdl.handle.net/10204/8114 en_ZA
dc.identifier.vancouvercitation Rosman BS, Ramamoorthy S. Action priors for learning domain invariances. 2015; http://hdl.handle.net/10204/8114. en_ZA
dc.identifier.ris TY - Article AU - Rosman, Benjamin S AU - Ramamoorthy, S AB - An agent tasked with solving a number of different decision making problems in similar environments has an opportunity to learn over a longer timescale than each individual task. Through examining solutions to different tasks, it can uncover behavioural invariances in the domain, by identifying actions to be prioritised in local contexts, invariant to task details. This information has the effect of greatly increasing the speed of solving new problems. We formalise this notion as action priors, defined as distributions over the action space, conditioned on environment state, and show how these can be learnt from a set of value functions. We apply action priors in the setting of reinforcement learning, to bias action selection during exploration. Aggressive use of action priors performs context based pruning of the available actions, thus reducing the complexity of look ahead during search. We additionally define action priors over observation features, rather than states, which provides further flexibility and generalisability, with the additional benefit of enabling feature selection. Action priors are demonstrated in experiments in a simulated factory environment and a large random graph domain, and show significant speed ups in learning new tasks. Furthermore, we argue that this mechanism is cognitively plausible, and is compatible with findings from cognitive psychology. DA - 2015-04 DB - ResearchSpace DP - CSIR KW - Search pruning KW - Action selection KW - Action ordering KW - Transfer learning KW - Reinforcement learning LK - https://researchspace.csir.co.za PY - 2015 SM - 1943-0604 T1 - Action priors for learning domain invariances TI - Action priors for learning domain invariances UR - http://hdl.handle.net/10204/8114 ER - en_ZA


Files in this item

This item appears in the following Collection(s)

Show simple item record