Agent-Based Modelling And Organisational Structure dekker/LA.ppt dekker/FINCX dekker/FINCY.

Agent-Based Modelling And Organisational Structure

http://www.acm.org/~dekker/LA.ppthttp://www.acm.org/~dekker/FINCXhttp://www.acm.org/~dekker/FINCYhttp://www.acm.org/~dekker/FINCZ

Dr. Tony Dekker ( [email protected] )

10 May 2002

2

Introduction

• The FINC methodology for analysing organisational structures

• Do FINC metrics predict organisational performance?

• Two simulation experiments

• What happened

• Java implementation: how we did it

• Intelligent agents & behaviour hierarchy

3

The FINC Methodology

IntN

• Force (Activity)• Intelligence (Information)• Network• C2 (Command and Control, Decision-Making)• Conceptual delays on network links need calibration

C2 C2

F F C2

F F

Int

N N

N

N

N

NN

4

The FINC Metrics

• Information flow coefficient (tempo superiority) low is good

= average path length (intelligence -> force)

• Coordination coefficient (coordination superiority) low is good

= average path length (force -> force)

• Intelligence coefficient (information superiority) high is good

= SUM (relevant area * (intelligence quality / path length))

Effective quality

5

The FINC Hypothesis

These metrics can predict organisational performance

i.e. better metrics mean the task gets done better

Test this using agent-based simulations

(later follow with real-world studies)

6

Experiment 1: the scenario

• A “SCUD Hunt”

• 4 SCUD missiles (white)

• Information from 1 satellite and 4 surveillance aircraft (green)

• Information of varying quality: “ghost”missiles (grey)

• 4 strike aircraft (blue)

• Several headquarters (red)

• 8 possible architectures

7

Experiment 1: the scenario

• A “SCUD Hunt”

• 4 SCUD missiles (white)

• Information from 1 satellite and 4 surveillance aircraft (green)

• Information of varying quality: “ghost”missiles (grey)

• 4 strike aircraft (blue)

• Several headquarters (red)

• 8 possible architectures

8

Experiment 1: the metrics

coordination coefficient (coord)

Hierarchical

Centralised

information flow coefficient (info)

intelligence coefficient (intel)

Hierarchical with Info Sharing

NegotiationCentralised with Info Sharing

Distributed

Negotiation with Info Sharing

Distributed with Info Sharing

9

Experiment 1: the results

Information superiority

is most important

Balance information & coordination superiority

Balance all three kinds of superiority

Balance information & tempo superiority

Poor Sensors Fair to Good Sensors

Slow Tempo

Moderate Tempo

Fast Tempo

10

Experiment 2: the scenario

• Hierarchical organisation of 16 companies

• Try to locate 5 “targets” and manoeuvre forces towards them

• World contains obstacles (green)

• 3 planning strategies

• 4 possible architectures

• Varying communications quality

11

Experiment 2: the results

95% of variance in performance predicted using only the intelligence coefficient

12

Implementation: Intelligent Agents & Messages

• Network of agents can be any graph

• Agents co-operate on the same goal

• Agents pass messages

• Agents have internal map and path planner:

Need support at (1,9)

Found target at (3,7)

I am at (4,5)

13

Implementation: Behaviour Hierarchy

Sim ple Long range

Sensor

Goto (point) Follow (agent)

Goal

Gun Radio jam m ing

W eapon

Behaviour

• Java Class Hierarchy

• Agents have “slots” for different behaviours

• Behaviour code manipulates agents internal map, etc.

14

Implementation: Dynamic Instantiation

• New Behaviour classes produced regularly

• Initial agent network from CAVALIER network editing tool

• CAVALIER specifies a text string (class name plus arguments) for each agent “slot”

• E.g. “Followgoal, A, B, C, D” means move towards average of A, B, C, D positions

• Agent simulation environment instantiates behaviour objects using dynamic object creation in Java’s reflection package

15

Summary

• FINC methodology for analysing organisational structures

• Do FINC metrics predict performance?

• Experiments 1 & 2

• Excellent prediction of performance in simulations

• Implementation: Intelligent Agents & Messages

• Behaviour Hierarchy & Dynamic Instantiation

16

Any Questions?

?