The New bayesian surprise rareness (BSR) measure for cognitive agents' reasoning and decision making
Loading...
Date
2008
Authors
Jeremiah, Davinna
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Bayesian Surprise is the study of surprise occurrence and as previously studied by others,
occurrences that were rare are considered surprising as well. However, through our study we
have found surprise and rareness to be conceptually different, thus both these notions have been
separated. Since the notions of surprise and rareness are now no longer considered the same,
we therefore have designed separate and new methods of computation for each notion and we
have named our method as Bayesian Surprise Rareness (BSR). Basically our computation is
based on the likelihood distribution unlike past works which measure surprise from the prior
and posterior distributions. The computation is based on the likelihood is due to the prior and
posterior distributions being not able to reflect every single occurrence of surprise and rare
cases. The concepts and computation of rareness and surprise are then applied to a multi-agent
system where the behaviours of agents are predicted. An agent that is being predicted may
exhibit rare behaviour and may give surprising responses which need to be identified and quantified.
Apart from that we also show the importance of the BSR measure in an agent's decision
making process through demonstrating its usefulness by using the result of the computation
to determine the optimal strategy so that accurate decisions can be made. We use influence
diagram to model our prediction and decision making process. However the existing influence
diagram is not able to reflect our computation methods and therefore we have extended it
to accommodate the surprise and rareness computations. The new graphical representation is
known as the Bayesian Surprise Rareness Incorporated Influence Diagram (BSRIID). We have
performed experiments to verify our computation methods and in one instant, decision making
for selecting a hint type using BSRIID is 14.48% more accurate then the traditional or standard
influence diagram. Then when the posterior is used to detect events that are surprising or rare,
only 9% are detected but as for the BSR measure, all events are detected.
Description
Keywords
Bayesian surprise rareness , Cognitive agents' , Decision making