Out zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA C. Upton - - PDF document

Using zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

Meta-Heuristics t o Explore the Many Would-be Worlds of Combat Simulations

Stephen zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

C.

Upton 18th ISMOR Oxford, UK zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

PmjectAlbe.t

@ MCCDC

27

• 31

August 2001

Out zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

I ine

TheIdea The Methodology Current Work Future Work Summary

1

The zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

Idea

continuous discrete coinplex stiucture interdependent nonsmooth noisy

. . .

chaotic nonlinear

. . .

• I. Find “interesting” regions zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

X b zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA C zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

X and YI: C Y

• 2. Find “relationships” among zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

x zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA EX-*, y E Yl:

Approach: apply optimum-seeking heuristics and data mining techniques

Generative Analysis

Proof

• of -Concept Project

The Idea

• Using [multi-agent] simulations and advanced(?) search heuristics

(optimum-seeking), have computer generate alternatives

. Rationale for idea

• Generate surprises
• Red-team Cat’s Pajamas concepts
• Support Future Concept Development and Exploration, e.g.,

Urban Warrior

• High Resolution - Low Fidelity - Urban setting
• What tactical concepts are most effective?

*

Test Idea - Simple Analysis Experiment

0 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

b

j zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

ect ive

T

• develop a methodology, and associated

software, that automatically finds and relates good solutions within the possibility space of a simulation, where "good" may be:

* the best or worst performing * the most robust (best performing over many environments)

.

the most interesting, e.g., the chaotic, the stable, the unstable

.

the most surprising or novel

Principal Components

• Algorithm(s) for generating a "solution"

Evaluation mechanism (object t o explore) - simulation

• r model

Algorithm(s) for analyzing a "solution"

• r set of

solutions Mechanism for generator-analyzer t o "talk"

• Representation for a potential "solution"

that can be manipulated by the generating algorithm(s)

* Specification o f fitness for a "solution" zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

3

I ' zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

Possible Generator zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

A laori

t

hms

Meta-heuristics Natural Algorithms

• Simulated Annealing, Evolutionary Algorithms,

Tabu Search Machine Learning A lgor

i

t zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

hms

Cultural Algorithms,. . .

. . . . .

Possible Analyzer Algorithms

Multiple Regression Cluster Analysis Association Analysis Bayesian Techniques Neural Networks

. . . . . . .

• .

E

• zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

I . q

WiylOnm.

• *

E

• 4

E

• W
• The Meta-Algorithm

S M T - zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

during "

A n a w

to M p

genefate next trial solutions zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

I-

Potential Uses

+Concept Exploration (e.g., tactics or system)

* Analysis o f Alternatives

• Force Structure Design

* Concept Generation

RedTeaming Simulation Verification zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

c

• Active Nonlinear

Test (ANT) idea

Validation (?)

Issues zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

• Estimating performance (fitness)

Runtime execution of simulation/model Ensuring diversity o f solutions

* Data structure operations (manipulating

general

* Multi-objectives * Role of constraints

. Ordering of search

• Knowledge extraction
• bjects)

_- zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

I

• Current Status

* Implemented basic GAn software framework (Condor,

JADE, NALEX)

* "Tested" software using JIVESTown with GA for

search on single machine (over rule base space)

* "Tested" software using Socrates with EP algorithm for

search on cluster of 20 machines (over weapon, comm, and sensor ranges)

* Implementing POSA, Cultured EP, Immune Algorithm * Adding code t o use with Mana zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

6

To Do List

Begin initial experiment with Cultured EP Post software as Albert-ware for general use Combine data mining algorithms, e.g., clustering, with heuristics t o possibly enhance search Combine statistical methodologies with heuristic techniques t o possibly enhance search Adapt generator operators for discrete combinations Examine multi-objective problems and generation of the Pareto optimal set Examine the role of constraints zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

Summary

We believe Generative Analysis is a new para and tool for the study and exploration of the possibility space o f a simulation simulations becomes feasible with advances n

• high-performance and distributed computing,
• automatic programming methods,
• heuristic search techniques, and
• agent-based modeling

The generation, and analysis, of a large number o f

Questions?

upton@mitre.org http://www.projectalbert.org/

References

aPr~;;~N~k#

Genetic Programming: On the Programming o f Computers by Means of Natural Selection, John R. Koza, 1992 Evolutionary Design by Computers, Peter J. Bentley (Editor), 1999 How t o Solve It: Modern Heuristics, Z. Michalewicz and D.B. Fogel, 2000 "Active Nonlinear Tests(ANTs) o f Complex Simulation Models", zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

J.

H.

Miller, Management Science, Vol. 44, No. 6, June 1998 "Multiojbective Optimization Usinq Evolutionary Alqorithms - A Comparative Case.Study", E. Zitzler and L. Thiele, FPSN V, LCNS 1498, pp 292-301,1998 "Distributed evolutionary algorithms f o r simulation optimization", H. Pierreval and J.L. Paris, IEEE Transactions on Systems Man and Cybernetics, Part A, Vol. 30(#1), pp 15-24, Jan 2000 "Unexpectedness as a measure o f interestingness in knowledge discovery", B. Padmanabhan and A. Tuzhilin", Decision Support Systems,

• Vol. 27. pp 303-318,1999

8