Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological...

24
Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander Hörnlein Christoph Oechslein Frank Puppe

Transcript of Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological...

Page 1: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary

processes

Alexander HörnleinChristoph Oechslein

Frank Puppe

Page 2: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

2 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Motivation / Problem

• Optimization of behavior in respect of– explicit evaluation function– implicit evaluation function

e.g. “the agents have to survive a certain period”

• Calibration towards a predefined target behaviore.g. “the agents should act exactly as in real life”

Page 3: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

3 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Evolution as optimization

• Population of potential solutions• Evaluation by means of “natural selection”• Iteration: Survivors (i.e. highly fit

individuals) reproduce

Page 4: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

4 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Reproduction

• Mutation– Offspring differs slightly - possibly advantageous– local search

• Recombination– Child possibly unites the advantages of both

parents– global search

Page 5: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

5 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Behavior in SeSAm

Agent

• Rules• Activities

• Parameters• Memory• Perception

IF (in activity1) ANDCondition THEN activity3

Activity1 Activity2

Activity3

Action1Action2

...

Page 6: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

6 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

GP approach: Mutation operators

activity

Parameter a += 10Approach agent xIncrease speed

Parameter a += 25Parameter b += 25Flee from agent x

Focus on earth

• Change numeric terminals• Change symbolic terminals

• Change non-terminals• Delete action• Add action

• Add new activity• Add new rule• Change rule• Delete activity• Delete rule

Page 7: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

7 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Advantage

• Extremely powerful• Little constraint by

initial structure of behavior

• Development of unnecessary or unwanted complexity

• Restrictions are difficult to define/set

• Slow• Hard to implement

within SeSAm

Disadvantages

Page 8: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

8 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

GA/ES approach: Mutation operators

activity

Parameter a += 10Approach agent xIncrease speed

• Change numeric terminals

Parameter a += 25 that’s it in principle.

Page 9: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

9 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Applicability of GA/ES approachwithin SeSAm

• Actions– Use numerical terminals– Can be controlled by probabilities

• Rules– Condition-parts use numerical terminals– Action-parts can be controlled by probabilities

Page 10: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

10 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Model modification

•Define rules for any reasonable transient

•Let evolution weight them

•Treat actions accordingly

Page 11: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

11 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Advantages

• Sufficient powerful• Easy to restrict:

Evolution can’t break boundaries of predefined behavior

• Fast• Implementation within

SeSAm is ‘straight-forward’

• Not extremely powerful

Disadvantage

Page 12: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

12 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

SeSAm genes

RULE: IF ENERGY > gene0 THEN MOVE

gene0:

(initial)value

(initial) standar

d deviatio

n

]upper

boundary

[lower

boundary

(initial) standard deviationdominancedistribution

(initial) value

lower boundaryupper boundary

Page 13: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

13 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

SeSAm genomes

agent role

behavior

family attribute

egg storage

genomedeclaration

gene0 declaration gene1 declaration ...

genome

gene0 gene1 ...allele0-0 allele1-0

gene0 allele0-1 gene1 allele1-1 ...... ...

Page 14: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

14 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Polyploid genome

• Treated threadwise • Treated genewise

dominancemutation

dominancemutation

Page 15: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

15 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

value0

value1

value2

meta gene

Possibilities for the gene-expression

•weighted

ii

iii

dominance

valuedominance

)(

)(

value0

•dominant/recessive

i

ivaluealleles#

1

•‘intermediary’

expression

Page 16: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Application

from individuals to colonies

Page 17: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

17 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Insects’ behavior

from ownreservoir

brood care

from nestreservoiridle

growfeed

feed on nestreservoir

feedon broodlay egg

mate

seeknew nest

seekmarker

setmarker insectsprey

hunt

fighttransportto nest

Page 18: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

18 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Insects’ genes

idle

grow

lay egg

mate

seek new nest

queen-factor

prey

hunt

fighttransportto nest

hunt-factor

from ownreservoir

brood care

from nestreservoir

brood care-factor

energy levelgenes

feed

feed on nestreservoir

feed on brood

egglevelgenes

Page 19: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

19 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Initial insects’ world

Page 20: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

20 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Insects’ world after 150,000 ticks

Page 21: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

21 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Changes of gene-pool

queen-factor

brood care-factor

hunt-factor

Page 22: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

22 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

More changes of gene-pool

initial egg energy

energy portion ant

energy portion brood

Page 23: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

23 / 23

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg

Optimization of simulated biological multi-agent systems by means of evolutionary processes

Results & Discussion

• Successful evaluation in three scenarios• ES/GA approach powerful and easy to use

? Use of explicit evaluation function for greater applicability

? Accelerate optimization (through parallelism)

Page 24: Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg Optimization of simulated biological multi-agent systems by means of evolutionary processes Alexander.

Lehrstuhl für Künstliche Intelligenz - Univ. Würzburg