GENERAL ENQUIRIES: Tel: + 27 12 841 2911 | Email: callcentre@csir.co.za

Show simple item record

dc.contributor.author Rosman, Benjamin S
dc.contributor.author Hawasly, M
dc.contributor.author Ramamoorthy, S
dc.date.accessioned 2017-05-16T10:19:45Z
dc.date.available 2017-05-16T10:19:45Z
dc.date.issued 2016-02
dc.identifier.citation Rosman, B.S, Hawasly, M. and Ramamoorthy, S. 2016. Bayesian policy reuse. Machine Learning, vol. 104(1): 99-127. DOI: 10.1007/s10994-016-5547-y en_US
dc.identifier.issn 0885-6125
dc.identifier.uri DOI: 10.1007/s10994-016-5547-y
dc.identifier.uri http://link.springer.com/article/10.1007/s10994-016-5547-y
dc.identifier.uri http://hdl.handle.net/10204/9043
dc.description © The Author(s) 2016. This is a pre-print version of the article. The definitive published version can be obtained from http://link.springer.com/article/10.1007/s10994-016-5547-y#enumeration en_US
dc.description.abstract A long-lived autonomous agent should be able to respond online to novel instances of tasks from a familiar domain. Acting online requires `fast' responses, in terms of rapid convergence, especially when the task instance has a short duration such as in applications involving interactions with humans. These requirements can be problematic for many established methods for learning to act. In domains where the agent knows that the task instance is drawn from a family of related tasks, albeit without access to the label of any given instance, it can choose to act through a process of policy reuse from a library in contrast to policy learning. In policy reuse, the agent has prior experience from the class of tasks in the form of a library of policies that were learnt from sample task instances during an offline training phase. We formalise the problem of policy reuse and present an algorithm for efficiently responding to a novel task instance by reusing a policy from this library of existing policies, where the choice is based on observed `signals' which correlate to policy performance. We achieve this by posing the problem as a Bayesian choice problem with a corresponding notion of an optimal response, but the computation of that response is in many cases intractable. Therefore, to reduce the computation cost of the posterior, we follow a Bayesian optimisation approach and define a set of policy selection functions, which balance exploration in the policy library against exploitation of previously tried policies, together with a model of expected performance of the policy library on their corresponding task instances. We validate our method in several simulated domains of interactive, short-duration episodic tasks, showing rapid convergence in unknown task variations. en_US
dc.description.sponsorship This research has benefitted from support by the UK Engineering and Physical Sciences Research Council (Grant Number EP/H012338/1) and the European Commission (TOMSY and SmartSociety grants). en_US
dc.language.iso en en_US
dc.publisher Springer Verlag en_US
dc.rights CC0 1.0 Universal *
dc.rights.uri http://creativecommons.org/publicdomain/zero/1.0/ *
dc.subject Policy Reuse en_US
dc.subject Reinforcement Learning en_US
dc.subject Online bandits en_US
dc.subject Transfer learning en_US
dc.subject Bayesian Optimisation en_US
dc.subject Bayesian Decision Theory en_US
dc.title Bayesian policy reuse en_US
dc.type Article en_US


Files in this item

The following license files are associated with this item:

This item appears in the following Collection(s)

Show simple item record

CC0 1.0 Universal Except where otherwise noted, this item's license is described as CC0 1.0 Universal

Search ResearchSpace


Advanced Search

Browse

My Account