[Colloq] Eric Robinson Thesis Defense TODAY

Patricia Freeman tricia at ccs.neu.edu
Tue Apr 22 09:38:42 EDT 2008


Eric Robinson will be defending his thesis TODAY at 1pm.
> The defense will take place in
> Room 366 WVH.
>
>
> Abstract:
>
> The problem of large implicit state space generation is examined.
> This has applications in areas such as artificial intelligence, where
> it can be used in puzzle search, group theory, where it can be used
> in orbit enumeration, and formal verification, where it can be used
> in explicit state model checking.  Cases of this problem that
> overflow distributed RAM are examined.
>
> Traditionally, two approaches are used for cases that overflow
> distributed RAM.  The first approach performs additional computation
> in order to allow for the state space to fit in RAM.  The second
> approach utilizes disk in a streaming manner in order to use it as an
> "extended RAM".  A hybrid approach is presented.  Our technique uses
> the available RAM to form an initial approximation of the state
> space.  It then goes to disk to discover where this approximation is
> incorrect.  We refer to this technique as "Tiered Duplicate
> Detection."
>
> With this and other techniques, a hierarchy of implicit state space
> generation techniques is built.  We refer to this as the "space-time
> search hierarchy."  In it, there is a trade-off between the amount of
> computation a technique must perform and the space required by that
> technique.  For each technique presented, we offer an analysis of the
> run-time and space requirements.  The analysis of the runtime takes
> into account the two main components discussed here, the number of
> computations along with the amount of data accessed on disk.
>
> Finally, we offer an API and corresponding software for the problem of
> large implicit state space generation.  Using this API, we implement
> our technique, Tiered Duplicate Detection, along with some sample
> applications.  With this, we demonstrate three things:  that the
> implementation of applications in our API is straightforward and
> fast; that our software can be used to effectively solve problems in
> large implicit state space generation; and that our analysis of the
> runtimes for these techniques are accurate when compared to the
> experimental times seen.



More information about the Colloq mailing list