Intention Reconsideration as Metareasoning (bibtex)
by Marc van Zee, Thomas Icard
Abstract:
We first present a theoretical framework for the intention reconsideration problem in MDPs, in the same spirit as much other work on metareasoning. This involves the construction of a meta-level MDP in which the two actions are ‘think’ or ‘act’. We then consider Kinny and Georgeff’s framework as a special case, reproducing their results, and comparing their agents to an “angelic” agent who decides optimally when to think or act. Interestingly, even the very simple agents Kinny and Georgeff considered behave nearly optimally in certain environments. However, no agent performs optimally across environments. Our results suggest that meta-meta-reasoning may indeed be called for in this setting, so that an agent might tune its reconsideration strategy flexibly to different environments.
Reference:
Intention Reconsideration as Metareasoning (Marc van Zee, Thomas Icard), In Bounded Optimality and Rational Metareasoning NIPS 2015 Workshop, 2015.
Bibtex Entry:
@InProceedings{vanzee-icard:borm2015,
  Title                    = {Intention Reconsideration as Metareasoning},
  Author                   = {Marc van Zee and Thomas Icard},
  Booktitle                = {Bounded Optimality and Rational Metareasoning NIPS 2015 Workshop},
  Year                     = {2015},
  Month                    = {December},
  Abstract                 = {We first present a theoretical framework for the intention reconsideration problem in MDPs, in the same spirit as much other work on metareasoning. This involves the construction of a meta-level MDP in which the two actions are ‘think’ or ‘act’. We then consider Kinny and Georgeff’s framework as a special case, reproducing their results, and comparing their agents to an “angelic” agent who decides optimally when to think or act. Interestingly, even the very simple agents Kinny and Georgeff considered behave nearly optimally in certain environments. However, no agent performs optimally across environments. Our results suggest that meta-meta-reasoning may indeed be called for in this setting, so that an agent might tune its reconsideration strategy flexibly to different environments.},
  Url                      = {http://www.marcvanzee.nl/publications/2015/borm2015_metareasoning.pdf}
}
Powered by bibtexbrowser