Transcript document

The Gold Seekers Project
Crescent Lab
Department of Computer Science
Texas Christian University
Three Sub-Projects

Alchemy



Gold Mind


A distributed process environment for AI applications
and cognitive models
Can run on a cluster of a grid
Cognitive modeling environment based on Alchemy
El Dorado


Improved distributed AI processing theory
Software/Hardware applications of Alchemy/Gold
Mind
Relationship of the Projects
Cognition
Concepts
Brain/Mind
Processing
Concepts
Goal Mind
Distributed
Processing
Concepts
Alchemy
New Concepts in
(Geographically)
Distributed
Computing
El Dorado
Useful
Applications
Improved Cognitive
Modeling Environment
Improved Special-Purpose
Distributed Environment
Alchemy
Geographically Distributed AI
System Requirements

Asynchronous processing nodes


Dynamic collections of nodes


most inference engines use some type of
constantly repeating match-select-fire cycle
dividing an application’s problem space into
sub-spaces requires a graph of reasoners
which share a flexible communication
protocol
Implementation needs to exhibit:



good overall processing times
speedup
secure communication
The Alchemy Approach

Build on other distributed architectures:



Create a GDAI friendly:





CORBA, IBM Aglets & Sun’s JINI
AMEBA (Hannon and Cook, 2001)
Threading model (housekeeping, clients, servers)
Security model (authentication, encryption, tracking)
Development environment (GUI-based)
Implement it using an OOD in C++
Test using different applications

Formally-defined web-based search of documents
Alchemy Components
SSL
Alchemy
Security
Model
Threading
Model
Alchemy
Client-Server
Model
Node
GADGET
Client
Server
Process
Control
Migration
Control
Handler
Housekeeper
Application
Nodes
Alchemy Architecture
GADGET
Migration Control
Client-Server Connections
Process Control
Node
Process Control
Process Control
Node
Node
Node
Node
Node
Node
Node
Processor Node
Process Control
Processor Node
Node
Processor Node
Node
Node
Processor Node
Alchemy Resources

The Crescent Lab Beowulf Cluster


Housed on the third floor TTC
12x2 + 7x1 processor array


28.4 GHz of total processing power
Uses 3 – switched 100baseT networks

Effective bandwidth of 600 Mb/sec
Handling Message Security

Authentication




Encryption




complex response
embedded time stamp
first key (public, super secure, pre-session)
algorithm (strong or weak)
public-private vs. symmetric keys
key and sieve length
Message tracking

digesting approach
Alchemy’s Approach

Embedded support


Multi-level



C++ class libraries based on SSL
support layer’s level is locked by system
application node level determined by server
Trade-off between speed and safety


256 bit Blowfish
MD5
Handling Network Security

Common port mapping



e.g., a local port to 8080 to local port
requires some kind of server interface and the application
intelligence to know what to route
VPN

to the inside network elements



to each other



looks like any other network device
has a IP address, mask and route, etc.
looks like a peer-to-peer (UDP) or client-server (TCP) connection
tunnels messages over a [secure] pipe
routes anything that knows the VPN’s address
Handling Network Security (2)

Intelligent Private Network (IPN)



Piggybacks on the Alchemy server
To simplify application design, Alchemy uses name spaces
 the application server’s port is maintained by the local
Alchemy server
 application clients always ask for server connections by
name, never by host and port
 the this name is translated into a host and port by the client’s
local Alchemy server and returned to the client via the
support system
Using IPN, when an application’s client ask for remote server, it
gets the host and port of the local IPN
(Picture of Beowulf here)
Gold Mind
The Gold Mind Mission



Develop a better understanding of human
cognition and interaction by modeling human
performance on a computer system
Support the creation of individual models of
humans performance on a given task which can
be later be integrated into larger models of more
complicated tasks
In the far distance future, solve the ‘grand
challenge’ problem of a generalized model of
human cognitions
The Gold Mind Architecture
The Etheron Computational Model
To Knowledge Tool
View/
Edit
System
Parent
Control
Process/
Inference
Agent
Parent
To Agent
(a parent
Etheron)
To Control
Daemon
Trace
To Trace Tool
Children
To System
(a parent
Etheron)
To other
Etherons
Listen for
connections
Current Gold Mind Models

TALLUS


STRESS


Explores a cognitive model for the Stroop effect
FAME


Explores language use and learning in young children
Explores a filter-fuser mechanism of attention and
arousal
ED-FAME

Explores how emotions control for attention and
arousal
The TALLUS Model

Teacher Assisted Language Learning and
Understanding Simulation



Contains three or more agents (one teacher and two
or more students)
Teacher is simply a HMI for entering adult level
speech and seeing the telegraphic output
Students have the same computational and cognitive
design, but different knowledge
The TALLUS Model Overview
One of the Model’s Student Agents
Social Knowledge Base
Ego Knowledge Base
Higher Order Processes Stimuli Router
A
g
e
n
t
R
o
u
t
e
r
Conversation Knowledge Base
Discourse Knowledge Base
Knowledge Stimuli Router
Semantic Concept Reasoner
Episodic Reasoner
Hearing
Interface
Utterance Stimuli Router
Semantics Reasoner
Syntactic Classifier
Lexicon Classifier
Vision
Interface
Vision Stimuli Router
Surface Structure Generator
Speech Stimuli Router
Vision Classifier
Speaking
Interface
Discourse Example from TALLUS
Teacher:
Jack:
Teacher:
Jill:
Teacher:
Jill:
Teacher:
Jack:
Teacher:
Jill, what is this? (holding a blue ball)
A ball.
Jack, let Jill answer. (pause) Jill, what is it?
A ball.
Yes, that is right! It is a ball. What color is it?
Blue ball.
Yes, that is right! It is a blue ball.
Jack, what do you do with it?
Play ball.
Yes, that is right! You can play with a ball.
TALLUS Speedup Results
6
100MB/sec
Speedup
5
500MB/sec
Linear
4
3
2
1
1
2
3
4
Number of Nodes
5
The STRESS Model

Stroop Test Response Evaluation SubSystem

Contains three agents (one researcher, one
evaluation and one subject)



Researcher agent is simply a HMI for entering testing
commands and viewing the subject’s responses
Evaluation agent generates test based on input from a
HMI
Subject agent is similar to a TALLUS student
The STRESS Model
Utterance
Generation
Reasoner
Task
Reasoner
HMI
Researcher
Agent
Concept
Reasoner
Utterance
Generation
SRN
HOP
SRN
Knowledge
SRN
Subjext
Agent
Utterance
Processing
SRN
Language /
Time Sync
I/F
Language
I/F
Language
I/F
Time Sync
I/F
Comm
Hub
Semantics
Reasoner
Color
Analyzer
Pattern
Analyzer
Syntax
Analyzer
Lexical
Analyzer
Time Sync
I/F
Stroop
Test
Generator
Vision
I/F
Vision
SRN
Vision
I/F
Evaluation
Agent
The STRESS Results
HUMAN
STRESS
Word Reading
130
Word Reading
Color Naming
120
Color Naming
800
Reaction time (milliseconds)
Reaction time (milliseconds)
900
700
600
500
110
100
90
80
70
60
50
400
40
Congruent
Control
Conflict
Congruent
Control
Conflict
The FAME Model

Attention and arousal function divided into



two functions – filtering and fusion
three aspects – time, space and modality
Resulting in 3 component types




Time Filter/Fuser (t)
Space Filter/Fuser (s)
Mixed Modality Filter/Fuser (m)
Allows these component to be combined as a
graph (most often a lattice) with sensor input and
processing elements
The FAME Application
t
Room
Temp #1
s
Temp #2
t
t
t
m
Light #1
s
Light #2
t
Attention
Control
Sensor
Processing
Arousal
Control
The ED-FAME Application
Model
El Dorado
Current El Dorado Projects

PRISE


LiMM


Integration of a mobile robot and room-based sensors
in a smart home environment to provide a mobile
assistant.
Multi-terrain sentry and inspection robot based on
biologically inspired design concepts.
RoAMS

Track-based arm assistant for assisted living
PRISE Environment
Mobile Sensor
and Tools Platform
Wireless
Network
Fixed
Sensors
Gold Mind Application
Alchemy
Crescent Beowulf Cluster
LiMM Environment
Sensor, Legs and
Manipulators
LAN/WAN Wireless
Network
Gold Mind Application
Gold Mind Application
Alchemy
Alchemy
The LiMM mController Network
Crescent Beowulf Cluster
The LiMM
The LiMM Robot Design Plan

Phase I




Build the robot platform and six legs
Uses 7 on-board µController for leg control
Power supplied by a tether
Phase II



Add end manipulators to the legs
Uses same set of on-board µControllers for
leg and manipulator control
Still uses tethered power
LiMM Design Plan (2)

Phase III



Add a sensor suite with its own set of
µControllers
Add an on-board power supply
Phase IV


Add an on-board µProcessor for remote
operations
Add a WAN network capability using a
satellite link or something like GSM
LiMM (Phase I) Facts

Each of the six legs contains



5 DC motors
A leg control module contain a rabbit µC
A Ethernet network connects the six leg
controllers to both a master controller the
the wireless network
RoAMS Environment
Fixed
Sensors
Gold Mind Application
Alchemy
Alchemy
The LiMM mController Network
Crescent Beowulf Cluster
Robot Arm and Track