Multi-Agent Explanation Strategies in
Real-Time Domains
Kumiko Tanaka-Ishii, Ian Frank
To be presented at ACL2000,
the 38th Annual Meeting of the Association for Computational Linguistics
Abstract:
We examine the benefits of using multiple agents to produce explanations
of complex domains. In particular, we focus on real-time domains such as
soccer commentary and car navigation. In such domains, having one agent
to deal with the tracking of any low-level events leaves the others free
to concentrate on the high-level picture. This simplifies the task of
each individual agent by allowing them to concentrate on a single
discourse strategy. Discourse issues such as controlling interruption,
abbreviation, and maintaining consistency are decomposed: rather than
considering them at the single level of one linear explanation they are
tackled separately within each individual agent and then also at the
level of inter-agent cooperation. We describe an actual implementation
of a system that uses this decomposition to produce real-time
multi-agent commentary for a game of simulated soccer. We evaluate this
system and show that it represents an advance over existing soccer
commentary programs.
|