IDA: A Cognitive Agent Architecture Lisa Soros. Outline Machine Consciousness Global Workspace Theory IDA: A Cognitive Agent Architecture

IDA: A Cognitive Agent Architecture

Lisa Soros

Outline

Machine Consciousness Global Workspace Theory IDA: A Cognitive Agent Architecture

Machine Consciousness


May 2001: Swartz Foundation sponsors a workshop at the Banbury Center in Long Island called 'Can A Machine Be Conscious?'

Attended by 20 psychologists, computer scientists, philosophers, physicists, neuroscientists, engineers, and indutrialists.

After the workshop, all but one attendee agreed that a machine could be conscious.


Assuming a machine could be conscious, how could we make it so?

First, we need a model of consciousness in general.


3 major theories of consciousness: 1. Dennett's Multiple Drafts Theory 2. Shanon's Theory 3. Baar's Global Workspace Theory

While each model has its idiosyncrasies, all assume that consciousness is not unitary in nature.

Global Workspace Theory


Theorized by Bernard Baars, a former Senior Fellow in Theoretical Neurobiology at The Neurosciences Institute (La Jolla, CA) in 1988

Emerged from the cognitive architecture tradition pioneered by Alan Newell and Herbert Simon (who won the Turing Award in 1975 for their contributions to AI and the psychology of human cognition)

Newell et al showed the utility of a global workspace in managing complex knowledge sources


According to GWT, The function of consciousness is to broadcast

information to separate functional modules throughout the brain

The global workspace is a central processor which contains the contents of consciousness

The global workspace functions as a sort of cognitive blackboard


Baars explains his theory using the metaphor of a theatre of consciousness:

Working memory provides the ”stage” of consciousness

Executive guidance directs the ”spotlight” of attention on the stage

The rest of the theater (small, special purpose, independent processes) is dark and unconscious



Developed by Stan Franklin and the 'Conscious' Software Research Group at the University of Memphis

Franklin's work on 'conscious' software agents has produced ~60 academic publications

After receiving his Ph.D. from UCLA, he has been on the faculty at UF, the Indian Institute of Technology at Kanpur, Carnegie Mellon, and the University of Memphis


IDA: Intelligent Distribution Agent Designed to serve as a detailer for the U.S.

Navy Communicates with sailors via natural language

e-mails to negotiate new assignments after their tours of duty

Must also adhere to ~90 Navy policies, fulfill job requirements, keep moving costs down, and respond to the wishes of the sailors



Codelet – a special purpose, relatvely independent mini-agent typically implemented as a small piece of code running as a separate thread (Hofstadter)

Perceptual codelets, attention codelets, information codelets, behavior codelets, language generation codelets...

Codelets in IDA correspond with processors in GWT



Perception Consists mostly of input from e-mails Uses only surface-level NLP (no parsing) Uses a Copycat-like architecture (Hofstadter &

Mitchell) with perceptual codelets triggered by both e-mail content and internal knowledge

This allows IDA to recognize, categorize, and understand


”A secretary sending out an email announcement of an upcoming seminar on Compact Operators on Banach Spaces can be said to have understood the organizer’s request that she do so even though she has no idea of what a Banach pace is much less what compact operators on them are. In most cases it would likely require person years of diligent effort to impart such knowledge. Nonetheless, the secretary understands the request at a level sufficient for her to get out the announcement. In the same way IDA understands incoming email messages well enough to do all the things she needs to with them.” - Franklin


Workspace Corresponds to the Working Memory (or STM) in

humans Consists of registers set aside for particular

categories of information Perceptual & internal codelets write to the

workspace, while many other codelets are watching the workspace in case they need to react to it

Interfaces with LTM; some, but not all, workspace contents are written to associative memory


Associative Memory SDM – Sparse Distributed Memory, a content-

addressable memory Writing to the workspace cues activation and

retrieval in the associative memory



'Consciousness' mechanism A coalition manager, a broadcast manager, a

spotlight controller, and attention codelets (which bring appropriate contents to 'consciousness')

Attention codelets watch out for items that might require 'conscious' attention

When attention-worthy items are found, the information codelets that describe the item are gathered into a coalition to handle the situation

The coalition then competes for the spotlight of 'consciousness'


Example: An attention codelet recognizes the please-find-job

message type It gathers information codelets carrying the sailor's

name and SSN in addition to the message type The attention and information codelets form a

coalition and compete for 'consciousness' If successful, the coalition's contents are broadcast

to the rest of the system


Behavior Selection (Decision Making) Behavior net: Several drives acting in parallel,

which vary in urgency with time and the environment

Behaviors depend on many behavior codelets Behaviors are like production rules, with

preconditions, additions, and deletions In the behavior net, behaviors spread activations to

other behaviors


Other modules: Emotion Deliberation Constraint Satisfaction Voluntary Action Negotiation Learning Metacognition

Documents

IDA: A Cognitive Agent Architecture Lisa Soros. Outline Machine Consciousness Global Workspace Theory IDA: A Cognitive Agent Architecture