microsoft powerpoint behaviourai new

Document Sample
microsoft powerpoint   behaviourai new Powered By Docstoc
					Behaviour-based AI in Distributed Complex Systems
Fang Wang Future Technologies Group Intelligent Systems Lab

Slide 2

Outline
• • • • • Knowledge-based AI Behaviour-based AI Agents Distributed complex systems Cell based optimisation in information ecosystems • Open problems

Slide 3

Knowledge-based AI
• Traditional AI
– Symbolic, knowledge-based and top-down control

• Focused on declarative knowledge structures
– Explicit knowledge representation and planning strategies

• The problem domain is usually static, predictable, closed and has no strict (time) constraints
– Indirect, well controlled interaction with the environment (the designer/operator acts as a mediator)

• Suitable for systems that demonstrate isolated and often advanced competences
– e.g., medical diagnosis and chess playing

Slide 4

Behaviour-based AI
• The real world is usually open, dynamic and uncertain, difficult to describe and requires real-time processing • Behaviour-based AI
– Situated, developmental, and bottom-up control
• Direct interaction with the environment and other entities • Improved internal structure over time based on past experiences
– Learning in real-time, changing environments

• Interaction dynamics among simple components that can lead to emergent complexity

System

input

output

Environment

Slide 5

Behaviour-based AI
• Autonomous, adaptive and robust
– How to decide what to do next for achieving a goal (goals)? – How to explore and update the internal structures and performance?

• Example techniques
– – – – Swarm intelligence Evolutionary Computation Artificial Neural Networks Autonomous agents

perception

action

states
past current next

input

output

Environment

• Reference
Maes, P. (1994) Modeling adaptive autonomous agents. Artificial Life: An Overview, pages 135-162.

Slide 6

Agents
• A computer system capable of acting independently, exhibiting control over their internal states (autonomous) • Flexibility
– Reactive – Proactive (goal-oriented) – Sociable (interactive)

• Research topics
– – – – BDI − Multi-agent planning − Coordination mechanisms Agent Communication Languages − Learning Matchmaking architectures and algorithms Negotiation strategies − Agent semantics & ontology …

• Reference
Luck, M., McBumey, P. and Preist, C. (2003) Agent Technology: Enabling Next Generation Computing. AgentLink

Slide 7

Distributed Complex Systems
• Properties
– Complex = Non*
• Nonlinear, nonstationary, nonreductionist, nonequilibrium, …

– Distributed, open, dynamic, uncertain, … – Heterogeneous, asynchronous, incomplete knowledge, …

• Require decentralised control, self-organising, adaptive and autonomous techniques • So as to be robust, scalable and resilient

Slide 8

Cell based optimisation in Information Ecosystems
• Information Ecosystems
– A complex web of information producers and consumers interacting in a constantly growing and changing environment
Database 22 Database
Agent Agent Agent Agent Agent Agent Agent Agent Agent Agent Agent Agent

Database 1 Database 1
Agent Agent Agent Agent Agent Agent

DB DB

DB DB

Agent Migration
Agent Agent Agent

Database 33 Database
Agent Agent

Database 44 Database
Agent Agent

DB DB

Agent Agent Agent Agent

DB DB

Slide 9

Problem
• Can a group of information processing agents collectively distribute themselves across a network to maximize the processing of user requests given the following:
– the agents only have local information, – the user demand patterns may change with time?

• Reference Rothermich, J., Wang, F. and Miller, J. (2003) Adaptivity in cell based optimization for information ecosystems. CEC03.

Slide 10

Cell based model
• Cell like agents
– Chemotaxis – Chemical signalling ─ Cell division

•

Cartesian Genetic Programming (CGP)
0. Do nothing 1. Move towards (2) 2. Move away from (2) 3. If (2) is present do (3), else do (4) 4. If (2) is present do (4), else do (3) 5. If energy is above UPPER_THRESHOLD do (3), else do (4) 6. If energy is below LOWER_THRESHOLD do (3), else do (4) 7. Perform (3) then divide 8. Perform (3) then release chemical signal

(2) External Input (3) Connection 1 (4) Connection 2 (1) Function Node number

0 1 2 3 4 5 |0000|7100|1010|1012|3100|3024|

0 0 0 0 0

1 0 0 7 1

0 1 0 1 2

0 1 2 1 3

1 0 0 3 4

0 2 4 3 5

Slide 11

Experiment Design
• Agents signal the number of untreated local requests as a chemical • The level of chemical at a single location = the number of outstanding local requests + the amount of chemical diffused from neighbor locations • Chemicals diffuse throughout the web of connected databases • Agents gain energy by processing user requests but lose energy continuously for other activities • The average number of new requests follows a Poisson distribution • The cells had 15% randomness in their direction of movement • Tests were conducted using 50 trials per experiment to test the performance of cell-based optimization

Slide 12

Change of population size in one experiment
• 10% more agents than the number of new requests • 87.83% of user requests completed
1700 1600

1500

1400

1300

1200

1100

1000

900

800 1 501 1001 1501 2001 2501 3001 3501 4001 4501 5001 5501 6001 6501 7001

T im e

A v e r a g e N u m b e r o f N e w R e q u e s ts P o p u la t io n S iz e

Slide 13

Evolution vs Human Design
• Human design cell program
If Energy > 30 Divide Else Move towards Food Chemical
AVERAGE % COMPLETE 93.74% 42.00% EVOLVED SOLUTION HUMAN DESIGNED SOLUTION

STANDARD DEVIATION

0.07

0.03

Slide 14

Adaptation
IMMEDIATE SHIFT GRADUAL SHIFT NO SHIFT (CONSTANT DEMAND)

Average Percent of Requests Completed
100% 95% 90%

AVERAGE % COMPLETE

87.83%

93.74%

81.08%

85% 80% 75% 70% 65%

STANDARD DEVIATION

0.11

0.07

0.11

60% 55% 50% 1000 2000 3000 4000 Time Immediate Shift Gradual Shift No Shift 5000 6000 7000

• Emergent behaviour
• Improved performance under gradual shift • Competition may have forced more diversity

Slide 15

Open problems
• Enhanced intelligence to deal with comprehensive and complex problems with multiple objectives
– Enhanced autonomy and adaptation for both individuals and groups – Application in large-scale distributed systems
• High robustness and scalability with less computational/communication cost

• Formal/high-level modelling and analysis
– To understand and apply technologies better


				
DOCUMENT INFO
Shared By:
Stats:
views:11
posted:12/18/2009
language:English
pages:8
Description: microsoft powerpoint behaviourai new