Department of Computer
Science
University of Essex
4pm Tuesday 14 July 1998
Room 2511, JCMB, King's Buildings
Dr Steel will be in the department on Tuesday only, hosted by Colin Stirling.
Standard artificial intelligence planning techniques (such as situation calculus) cannot handle plans where some of the steps are not actions on the world (like "buy bread" or "go to London") but actions affecting the plan actually being executed (such as "choose what to buy" or "plan how to get to London").
Operational semantics seems to provide an alternative. One can at least propose that the agent follows CCS-like rules that can be caricatured as
Plan is a plan that achieves Goal
------------------------------------------------------
make plan for Goal
make plan for Goal; do it ----------------------> Plan
However, while trying to use operational semantics like this I have
come to see a closer relation between it and possible world
accounts. Suppose possible histories --- sequences of possible
worlds --- are defined by the sequence of the agent's "mental
states" as well as its actions. An observer looks on. As time
advances the observer (generally) has better and better
information, and some possible histories are rendered incredible.
If one identifies histories that the observer cannot distinguish,
the "parallel" possible histories group tree-like shapes apparently
identical with the synchronization trees of operational semantics.
It seems, in fact, that there may be a systematic way of understanding process calculus as an abstraction of of sets of possible histories that an observer cannot distinguish.
This seminar is an attempt to get feedback about whether this observation is either novel or useful. (I should say that I believe that, though this proposal may be misguided, it is not a re-invention of Hennessy-Milner logic: which the abstract might suggest.)
| Other LFCS Theory Seminars |
Ian Stark Monday 15 June 1998 |