consortiums

PROJECT 7

DISTRIBUTION AND ASSIGNMENT

Stanley P. Franklin
University of Memphis

Proposed Grant $59,596

Abstract:
The roughly three hundred detailers represent some one hundred different communities, each comprising a group of sailors with common job skills and pay grades. The detailer(s) for each community (say sonar technician chiefs) is chosen from that community in order that he or she possesses the required domain knowledge relating to that community. As a human detailer does, the IDA (Intelligent

Distribution Agent) prototype will be designed to work within a single community. This will require the gathering and encoding of knowledge (knowledge engineering) to two kinds, that common to all detailers, and that specific to a single community.

The IDA architecture implements the global workspace theory of consciousness (Baars 1988, 1997). Though it’s a multi-agent system with the work being done by individual small codelets, it utilizes modules devoted to perception (natural language processing), associative and episodic memory, emotions, "consciousness," action selection, constraint satisfaction, deliberation, negotiation (natural language generation), learning and metacognition. The implementation of many of these modules are motivated by mechanisms from the "new AI" such as sparse distributed memory (Kanerva 1988), behavior nets (Maes 1989), the Copycat architecuture (Hofstadter & Mitchell 1994), and pandemonium theory (Jackson 1987). Most modules utilize a workspace (working memory) containing registers for data structures. Natural language understanding and generation are done using surface features. The "consciousness" mechanism and the behavior net play the central role in IDA activities.

Here we propose to plan and design a development period for sailor agents and for command agents. During this period, such an agent would systematically learn the needed domain detailer knowledge from a human detailer. This would be accomplished in stages. At first the agent would watch the human detailer asking questions about concepts not understood. Later the agent would work under the supervision of the human detailer on cases designed for training. Finally the agent would work unsupervised.