IDA, a Software Agent Cognitive System
by Stan Franklin
A US Navy sailor writes to IDA, "I'm approaching the end of this tour of duty and will need a new job. Can you find me something suitable in the San Diego area?" A couple of months of negotiation via a dozen or so email messages follow. "Can I have the job at the El Toro Marine Base?" "No, you must go to sea duty, and that temporary detached duty won't count." "Can I wait two weeks for the next requisition [jobs] list." "Sure, maybe something more suitable will come up." Issues concerning the sailor's preferences, job qualifications, and Navy policy are discussed. Finally the sailor agrees to an offered job aboard the USS Pennington based out of San Diego. A negotiation between a sailor and a human personnel officer? No, between a sailor and IDA, a software agent cognitive system.
IDA's environment consists of a typical fast workstation with lots of memory, a connection to the Internet, and access to several US Navy databases for personnel records, training schedules, etc. She senses her environment by interpreting incoming strings of symbols from email messages and database records, and acts on it by composing and sending her own such email messages and database queries. She pursues her own agenda, all of which has to do with persuading a sailor to accept a job she has selected as suitable and offered. IDA is a software agent that goes a long way towards being a cognitive system. Her job is to completely replace a human 'detailer,' the personnel officer who normally performs such negotiations.
Supported by the US Navy over the past five years, the IDA project is the work of the 'Conscious' Software Research Group at the University of Memphis. It pursues two objectives: 1) to automate the Navy's assignment of personnel, and 2) to model human cognition in the process. 'Conscious' in the sense that she models a psychological theory of consciousness, Baars' global workspace theory, IDA also models theories of perception, memory, decision making, deliberation and constraint satisfaction. She is a fertile source of hypotheses about human cognition that can be tested by cognitive scientists and neuroscientists.
The same IDA technology is capable of automating the tasks of many human information agents, that is people who negotiate in natural language, access information in databases, adhere to company or agency policy, make decisions, and create paper products. Such human information agents might include customer services agents, travel agents, loan officers for a bank, and various low to mid-level public servants. Though the IDA technology would be common to all of these, a considerable chore of gathering and encoding appropriate knowledge would be required to implement any one of these.
This IDA technology is based on a host of disparate mechanisms taken from the 'new' artificial intelligence. These include the Hofstadter and Mitchell's Copycat architecture, Kanerva's sparse distributed memory, Maes' behavior nets, and Jackson's pandemonium theory. IDA is currently up and running, and has been tested to the satisfaction of Navy detailers. Watching IDA in action, their reaction is typically a nod of the head together with "Yes, that's how I do it."
This IDA technology is currently being adapted to produce a web-based multi-agent system in which each of the more than 300,000 sailors will be represented by his or her own IDA-like agent devoted to the sailor's interest. The sailor will communicate with his or her agent in a private chat room environment on the web. Each of the Navy's commands will also be represented by an IDA-like agent looking out for its needs. Between periodic consultations with its sailor, a sailor agent will negotiate with various command agents until some agreement is reached. If both the sailor and the command approve the agreement made by their agents, it's taken before an IDA-like broker agent who is concerned with adherence to Navy personnel policies. The broker agent's approval constitutes the final blessing on the agreement. The sailor will then receive orders to the agreed upon job.
If a multi-agent system consisting of hundreds of thousands of quite sophisticated software agents seems impossible, that's because it is impossible. To make this project feasible, we partition the system along the Navy's community structure. Each community consists of 500 to 5000 sailors of similar paygrade and job skills. One such community might be that of sonar technicians rated E7 through E9. A separate multi-agent job assignment system based on the IDA technology will be developed for each such community.
On the science side of the project we hope to implement a number of cognitive features of the IDA conceptual model that are not currently part of the running computational model. These include a transient episodic memory, the automization of action sequences, the involvement of emotions in decision-making and problem solving with novel problem. Mechanisms for all these have been designed and await resources for coding and testing.
Much additional information about the IDA project, including many technical details can be found at the link below in the form of online versions of published papers.
Stan Franklin, The University of Memphis, USA