Intention without Representation

Peter Wallis

Friday, 28th May 2004
0930 - 1100

Theatre 2
University of Melbourne ICT Building
111 Barry Street, Carlton.

Abstract:

Representation is a central issue in classic AI. In the late 80's and 90's there was considerable interest in solving the representation problem by avoiding it and using not-representational approaches to classic AI tasks. Some claimed that neural nets had a distributed representation that in some way avoided the symbol grounding problem. Others such as Rodney Brooks at MIT simply hard-wired behaviour into situated agents. This talk is based on a paper that is to appear in Philosophical Psychology that describes how planning, in the full means-ends sense, can be achieved without representing the world in any way. The mechanism is based on BDI, but uses a plan library containing data structures called Goal Tagged Activities. While the mechanism does not solve the symbol grounding problem, it does push back the boundary at which rational behaviour requires symbols. The mechanism described enables a range of applications that require more than insect level intelligence and the presentation finishes with a discussion of teamwork.

Biosketch:

Peter Wallis has a PhD from RMIT on semantics for search engines and has been active in the Natural Language Processing community in Australia since 1989. From 1995 to 2001, he worked primarily for Defence on information extraction from text, and on conversational agents. While at Defence he was studying for an MBA and currently does consulting through his own company.