Agent-Ba Base sed Sim imula lation ion Dagstuhl Seminar Modeling - - PowerPoint PPT Presentation

▶
agent ba base sed sim imula lation ion
SMART_READER_LITE
LIVE PREVIEW

Agent-Ba Base sed Sim imula lation ion Dagstuhl Seminar Modeling - - PowerPoint PPT Presentation

Oct 29, 2023 135 likes •496 views

Agent-Ba Base sed Sim imula lation ion Dagstuhl Seminar Modeling and Analysis of Semiconductor Supply Chains February 10 th , 2016 Dr. Iris is Lorsc rsche heid id Hamburg burg University iversity of Techn hnology, ology, Germ

slide-1
SLIDE 1

1

Agent-Ba Base sed Sim imula lation ion

Dagstuhl Seminar „Modeling and Analysis of Semiconductor Supply Chains“ February 10th, 2016

  • Dr. Iris

is Lorsc rsche heid id Hamburg burg University iversity of Techn hnology,

  • logy, Germ

rmany any www.tuhh. tuhh.de de/mac /maccs cs - Institute titute of Mana nage gemen ent t Accounting counting and Sim imula lation tion

slide-2
SLIDE 2

2

A few words about myself… Current research

– Using Structural al Equation

  • n Models (SEM-PLS)

LS) to emp mpirical ally y valida date e agent architec ectures res (see, e.g., Lorscheid et al. 2014) and to analyz yze simulated d data (see, e.g., Mertens et al. 2015). – Developing standards rds for the design n of simulation n experime riment nts and analys ysis s of (complex) simulation models (see, e.g., Lorscheid et al. 2012, Lorscheid and Meyer in press). – Learni ning g agents for human complex systems (see, e.g., Lorscheid 2014).

Community

– European Social Simulation Association (ESSA)

http://www.essa.eu.org

slide-3
SLIDE 3

3

Perspect ctive ives s on agent-base sed sim imula lation ion and it its contr tribu ibution tion Analyzing Emergence Incorporating human behavior Using artificial agents

slide-4
SLIDE 4

4

About emergence…

  • https://www.youtube.com/watch?v=KGeg

mMWsgu8

slide-5
SLIDE 5

5

Fir irst step: Underst rstand anding ing in indiv ividu idual l behavio vior Decision rules:

  • 1. Separation - avoid crowding neighbors (short range repulsion)
  • 2. Alignment - steer towards average heading of neighbors
  • 3. Cohesion - steer towards average position of neighbors (long range

attraction)

slide-6
SLIDE 6

6

Sim imula lation: ion: Flo lockin king behavior vior

slide-7
SLIDE 7

7

Syste tem m Behavio avior r is is t the Result lt of Indiv ividu idual l Actio ion

Agent behavio avior Macr cro level vel Micr cro level vel

Behavior rules Preferences Perceptions (Information) Communication Interaction of individuals

System success? Individual Strategies

slide-8
SLIDE 8

8

„If you grow the phenomena you‘ll understand how it works.“ Joshua M. Epstein Through the bottom-up design, self lf-org

  • rganiza

nization tion and emergent nt processes sses may evolve that are not explic licit itly ly modelled lled (or even understood by the modeler!) Macal

and North 2010

slide-9
SLIDE 9

9

Perspect ctive ives s on agent-base sed sim imula lation ion and it its contr tribu ibution tion  Analyzing Emergence Incor

  • rpor
  • rat

ating ing huma man n be behavio vior Using artificial agents

slide-10
SLIDE 10

10

Agent

  • Representing the decision rules of the

agent explicitly

  • Representing the interdependencies

between the different human activities

  • Modelling the individual heterogeneity
  • f agents

social re-/proactive autonom. adaptive Environment Action Perception Agent Agent attributes Decision rules, Preferences, Perceptions

slide-11
SLIDE 11

11

What are the effect cts s of varying ying decis isio ion models? ls?

  • Approaching more complex („human“) behavior models…
  • Perfect / rational

Agents Human decision makers

  • decision maker
  • (homo oeconomicus)
  • … to explore possible consequences

Preferences, Information, Condition -> Action Rules etc.

slide-12
SLIDE 12

12

How w do we in inform m the decis isio ion n makers s wit ith ABM?

  • 1. ABM asks new questions, like:

– What are the individual strategies? – What is the „human factor“ in your processes? – What are the interactions?

  • 2. Implementation of findings in an ABM to analyze the

effect.

– How big is the difference that it makes? – How robust ist the system? – What are better ways of designing your environment? ABM is about understa standing nding the effects of human behavior to change (vs. tool to support local decisions)

Empirical study Modeling & Simulation study

slide-13
SLIDE 13

13

Under erst stand and huma man n behavior vior to design ign your ur envir iron

  • nme

ment! nt!

slide-14
SLIDE 14

14

Under erst stand and huma man n behavior vior to design ign your ur envir iron

  • nme

ment! nt! „As well as you need to make sure that people escape safely, you need to give them tools to make sure that they not get stucked in planning.“

  • 1. Understand human (re-)actions.
  • 2. Design your walls and columns!
slide-15
SLIDE 15

15

How to unde derstand stand human behavior?

slide-16
SLIDE 16

16

How w to understa stand nd human man behavior vior?

  • Surveys, interviews, lab experiments
  • Using theories, such as

– Prospect theory (Kahnemann/Tversky 1979) – Learning algorithms – See: Bounded rationality / behavioral economics

  • New ways, e.g.:

– Using structural equation models for agent modeling (Lorscheid et al. 2014: The PLS Agent - Agent Behavior Validation by Partial Least

  • Squares. Social Simulation Conference)
slide-17
SLIDE 17

17

Examp mple le from m the Infin ineo eon n project: ct: Supply ly chain in pla lanner

Field Research Personal Experience Qualitative Research General IFX Information Flow Work Shadowing Interviews

(Blumberg and Atre (2003) estimate that around 85% of business information does not exist in a structured way)

slide-18
SLIDE 18

18

The in influ luencing cing facts s of a d decis isio ion (Infine fineon

  • n)

What are central aspects of planning, affected by human decisions?

  • Where are individual

deviations?

  • What are the difficult

situations? (beyond pressing the button)

  • How do newcomers learn

to act?

  • (…)

Decision Behavior

  • f a SCP

IFX Planning Processes Planning Tools Planning Reports Rules/ Guidelines/ Targets Personal Experience Business Environ- ment Product Specifics Work Environ- ment

slide-19
SLIDE 19

19

Dir irect in interactions tions of the pla lannin ing agent

slide-20
SLIDE 20

20

Data collec llectio ion n research ch design ign

Information Content

Scientific Literature Qualitative Research Infineon Processes Conceptual Model Modelling & Simulation

Continuous development & improvement of the current IFX processes

Level of Detail

Königer 2016

slide-21
SLIDE 21

21

How to mode del human behavior?

slide-22
SLIDE 22

22

Sim imple le agent archit itectur cture

Agent Environment S A a x Rules f(x) = a

slide-23
SLIDE 23

23

Knowle

  • wledgebase

ase

slide-24
SLIDE 24

24

Learning ing as addit itio ional l featur ure e to in increase se autonomy

  • nomy

x state of the environment y new enviornmental state z hidden factors influencing the

  • utcome of the agents’ decision

r reward a agents‘ action

slide-25
SLIDE 25

25

Category egory Divisi sion Manager er Learn arning model el requireme ements (A) K Knowledge

  • True productivity value
  • Interval of possible

productivity values

  • Resources
  • Compensation value
  • Reported productivity of
  • pponent

Values in knowledgebase [R1.a]

  • Prediction of opponent’s

behavior Belief [R.1b] (B) Prefer ference ces Maximizing the compensation Profit maximization [R2] (C) Reaso asoning strategy Report with highest expected individual payment Best expected individual utility [R3] (D) Experie rience ce collect ction

  • Value of strategies

Implicit experience collection[R4.a]

  • History about past

experience Explicit experience collection [R4.b] (E) E Exploration strategy

  • Random report

Random exploration [R5.a]

  • Systematic exploration of

under-, overestimation, and truthful reporting Systematic exploration [R5.b] (F) Pay ayoff ff Incentive scheme defines positive payments only. Positive payoff [R6]

slide-26
SLIDE 26

26

Behavior models

Behavior Model Learning Algorithm Knowledge base Random Zero Intelligence #,# → report BehM1 Sarin Vahid #,# → report BehM2 Sarin Vahid productivity, # → report BehM3 Ficticious Play productivity, prediction → report Rational

slide-27
SLIDE 27

27

Behavior models

Behavior Model Learning Algorithm Knowledge base Random Zero Intelligence #,# → report BehM1 Sarin Vahid #,# → report BehM2 Sarin Vahid productivity, # → report BehM3 Ficticious Play productivity, prediction → report Rational

slide-28
SLIDE 28

28

Results under varying incentives

Effect of adapted incentives

slide-29
SLIDE 29

29

Perspect ctive ives s on agent-base sed sim imula lation ion and it its contr tribu ibution tion  Analyzing Emergence  Incorporating human behavior

 How to understand human behavior?  How to model human behavior?

Usin ing artif ific icia ial l agents

slide-30
SLIDE 30

30

The variety of agents…

  • Agent may vary from very simple condition-action patterns to

complex, intelligents entities. Agent as „Softbot“ Robot without a body

slide-31
SLIDE 31

31

Artif ific icia ial l Intellige lligence ce

Russel and Norvig 2009, p.2

slide-32
SLIDE 32

32

Agents s may fin ind good solu lutions! ions!

slide-33
SLIDE 33

33

Agents s may fin ind good solu lutions! ions!

slide-34
SLIDE 34

34

Perspect ctive ives s on agent-base sed sim imula lation ion and it its contr tribu ibution tion  Analyzing Emergence  Incorporating human behavior

 How to understand human behavior?  How to model human behavior?

 Using artificial agents

slide-35
SLIDE 35

35

Finally…

Incorporate human behavior in your models: „…this complexity exists, if you wish it away or not!"

(Ken, in a different context)

Reader: ;-)