Lucio Biggiero 2009

Lo studio dei sistemi complessi attraverso la simulazione ad

agenti interagenti:prospettive applicative nelle scienze

sociali e in ecologia

Lucio BiggieroUniversità dell’Aquila, Knownetlab Research Center

www.knownetlab.it, lucio.biggiero@gmail.com

Lucio Biggiero 2009

An overview of ABSM

• 1) What ABSM are

• 2) Some epistemological and methodological aspect

• 3) Some categorization of simulation methods

• 4) A categorization and example of ABSM

• 5) Some suggestions for future works

Lucio Biggiero 2009

1) What ABSM are• Artificial laboratories for generative

experiments

• Computer programs, that is more or less complex algorithms

• They are more than tools: they are theories/models

• They are experimental laboratories that can provide social scientists with the same tools of natural scientists

Lucio Biggiero 2009

A historical perspective

Artificialintelligence(50’-60’)

Artificial life(70’-…)

Artificialsocieties(80’-…)

Laboratories for

Computerscience

biology Socialsciences

Lucio Biggiero 2009

General Traits of ABSM

• They can be theory laden

• They can be data laden

• They can be empirically testable

• ABSM Are Theories in Action into the Virtual World

• ABSM lead social sciences from dogma and doxa to episteme

Lucio Biggiero 2009

They “reproduce” reality

• In principle, they have no limit to reproducibility

• In practice, every model should be purposefully built, and its complexity seriously hinders its viability

• Its algorithmic nature guarantees for the internal consistency

Lucio Biggiero 2009

The principle of generative explanation (Epstein, 2006)

• If (1) the model of a given macro-phenomenon can be implemented through plausible theoretical hypotheses concerning agents, their interactions and environment;

• If (2) there is likelihood between the simulation outcomes and the empirical findings;

• Then those theoretical hypotheses can be considered as sufficient conditions to explain that phenomenon

Lucio Biggiero 2009

2) Some epistemological and methodological aspect

Lucio Biggiero 2009

Properties of social phenomena

• Because of very high phenomenological complexity, they require a huge size of empirical base necessary to build or test theories

• Phenomenological complexity: instability, roles of expectations, emergent properties, sensitivity to interactions between observer-observed, path dependency, nonlinearity, high interconnectedness between factors

Lucio Biggiero 2009

Which kind of phenomena are better suited for ABSM?

• Many complex agents interacting in complex ways: the micro-macro issue

• Autonomous cognitive agents• Agents characterized by multiple

(even contradictory) behaviors (and hence their preferences should not be forcedly described by expected subjective utility functions)

• Equilibrium is not a must

Lucio Biggiero 2009

Human traits of agents

1. Perception-distinction capability;

2. Intentionality;

3. Goal-seeking;

4. Memory (information storing);

5. Heterogeneity;

6. Communication;

7. Rule creation-following;

8. Cheating.

Lucio Biggiero 2009

Social complexity• Emphasis on interactions respect to

matters

• Emphasis on complex and often unexpected outcomes produced by simple rules

• Reduced role for collective minds (deliberated social rules)

• Weakening of the links between intentionality and its outcomes

Lucio Biggiero 2009

Ideal methodological process

Current knowledge

Simulation model(in virtuo)

Field research

(in vivo)

Laboratory cases(in vitro)

Lucio Biggiero 2009

3. Some categorization of various types of simulation models

• top-down vs. bottom-up (emergence)

• static vs. dynamic

• function-based vs. system-based

• many low complex parts vs. few highly complex parts

Lucio Biggiero 2009

Examples • Econometric models• Structural equation models• Network analysis• Neural nets• Cellular automata• NK-FL• Game theories• System dynamics• Agent-based models

Lucio Biggiero 2009

Which Kind of Simulations?An Epistemological View

• Inner logical structure: once described agents’ behavioral characteristics, what should we expect they did or will do after n iterations (interactions)?

• Depending on the model purpose, structure and validation chance, its outcome can range from scenario analysis to testable predictions and retro-dictions

Lucio Biggiero 2009

In practice

• If we know the goal, we use models to search for the better patterns (behaviors) to reach it

• If we don’t know the goal, but we know the current patterns, we use models to study long term evolution

• If we know both, we can use models to explore what happens with different goals or different patterns or both

Lucio Biggiero 2009

4. Some categorization and example of ABSM

Lucio Biggiero 2009

Categorizations • Three broad categories of in terms of level

of abstraction/generalization:

• 1) case studies;

• 2) middle range models;

• 3) abstract models.

• First- and second-order emergentist

• Merely rules applying or even creating

• With or without (and levels of) learning

Lucio Biggiero 2009

Case studies

• Modeling a specific entity• The aerospace industrial cluster of Rome

in the 2005• The FIAT case in 2008• Etc.• Understanding Anasazi culture change

through agent-based modeling• Dean et al., 2000. in Kohler & Gumerman (eds.) Dynamics in human

and primates societies: agent-based modeling for social and spatial processes. Oxford: Oxford UP.

Lucio Biggiero 2009

Anasazi culture change

• Ethnic group living between 1800 and 1200 BC

• They suddenly disappeared: why?

• The traditional explanation addressed to climate changes

• Conversely, the model suggests social-political factors

Lucio Biggiero 2009

Abstract models

• They point at very general issues, which are common to many fields and/or do not depend significantly on specific circumstances

• Example: the emergence of cooperation between people, insects, virus, molecules. See literature on direct and indirect reciprocity, the evolution of cooperation, etc.

Lucio Biggiero 2009

Schelling’s model (1971, 1978) on racial segregation in American cities• Black and white agents are randomly

placed on a grid whose cells represent filled or empty households

• For levels of the tolerance threshold at or above 0.3, an initially random distribution of households segregates into patches of black and white, with households of each color clustering together

Lucio Biggiero 2009

Opinion dynamics orthe fragility of democracy

How can opinions, which are initially considered as extreme and marginal, manage to become the norm in large

parts of a population?Deffuant et al., 2002

i.e. the Nazis, the Bolshevists, the Maoists, the radical Islamists, the ecologists?, etc. growth and dominance

Lucio Biggiero 2009

A simple model structure•Agents have an opinion (a real number between -1 and +1) with a certain degree of uncertainty and interact randomly•Agent j is affected by the opinion of agent i by an amount proportional to the difference between their opinions, multiplied by the amount of overlap divided by agent i’s uncertainty, minus 1•Excepted few extremists with most positive or negative certain opinions, most agents start with an opinion taken from a uniform random distribution and with a common level of uncertainty

Lucio Biggiero 2009

Outcomes and predictions

• Under these conditions extremism spreads leading the population towards one of the two opposite

extremes

• Without extremists, the population would converge on the moderate

opinions

Lucio Biggiero 2009

• Abstract models

• The COD (Coordination for Organization Design) Model (Biggiero & Sevi, 2009)

• Middle range models

• The CIOPS (Cognitive Inter-organizational Production System) Model (Biggiero & Sevi, 2009)

• The KNOWTIC (Knowledge Transfer within Industrial Clusters) Model (Biggiero & Basevi, 2009)

Lucio Biggiero 2009

The COD model

Task interdependencies

Parallel

Sequential

Reciprocal

Emergent effects of task interdependence and bounded rationality on workgroup performance

Lucio Biggiero 2009

Coordination

Looking and engaging

Agent

Agent

Agent

The simulation model

The model structure

Lucio Biggiero 2009

Main results

simple people perform better when coordinated by simple rules

Computational capacity

low high

high

low

Coordination complexity

Effective combination amongBounded rationality and coordination complexity

effective

effective

ineffective

ineffective

Lucio Biggiero 2009

…main results we obtained…

1. an algorithmic confirmation of the law of requisite variety;

2. an algorithmic confirmation of the ordering of interdependencies in terms of complexity;

3. an algorithmic confirmation of the fit between task interdependencies and coordination mechanisms;

4. a formalization of task interdependencies and bounded rationality in terms of computational capacity;

5. an algorithmic analysis of the combined effects of bounded rationality, task interdependencies and coordination mechanisms on workgroup performance

Lucio Biggiero 2009

Selection Devices1. Random Choice2. Direct Experience3. Indirect Experience4. Reputation

Based onInformation transfer

Profits ofclients

Quality ofpurchases

Suppliers’quality

Goal of clients: choosing thebest suppliers

Opportunism by cheating and its effects on industry profitability

Lucio Biggiero 2009

Filiere and market structure

market 1: downstream / intermediate

Producer A

Producer B

Producer C

Market 2: intermediate / upstream

Supplier 1

Supplier 2

Supplier 3

Producer 1

Producer 2

Producer 3

Supplier α

Supplier β

Supplier γ

Sequential Technology

segment 1: downstream segment 2: intermediate

segment 3: upstream

Producer D Supplier δSupplier 4 Producer 4

Lucio Biggiero 2009

n5

n3

n4

n2

n4n1

n4n2n4n3

n3n1

n3n2

n3n5

n4 cognitive network

n3 cognitive network

n4 perception of n1Attributes:- Quality- Reliability as informer

sources:- direct experience-based trust- indirect experience-based trust- reputation-based trust

Business relationshipsIndustry n1

Interactions between the structural and the cognitive networks

Lucio Biggiero 2009

100

130

160

190

220

250

280

310

0 100 200 300 400

RND

DEBT

INDEBT.0

REBT.0

DEBT produces a worse performance than INDEBT and REBT

REBT ensures more stability and higher average profit than INDEBT

reliable communication makes easier and faster the information space

exploration.

Honest final producers – general results

Lucio Biggiero 2009

When all agents are full cheaters (REBT.1), profitability oscillates around a profit

that is near, and sometimes below, that produced by RND

when only false information are shared and firms rely on reputation suggestions,

agents cognitive efforts are totally wasted and the worst performance is observed.

Final producers – Cheating effects on REBT

100

130

160

190

220

250

280

310

0 100 200 300 400

REBT.0

REBT.50

REBT.1

Rebt.0: agents do not cheatRebt.50: agents cheat half timesRebt.1: agents always cheat

Lucio Biggiero 2009

When agents cheat, it is better to trust direct than indirect experience in order to avoid

false information.

Information reliability (quality) is more strategic than its quantity.

Effectiveness of decision making patterns

Cheating agentsCheating agentsHonest

agents

Honest

agents

Eff

ect

iven

ess

Eff

ect

iven

ess REBT

INDEBT

DEBT

DEBT

INDEBT

REBT

High

Low

High

Low

Lucio Biggiero 2009

Even if they are submitted to the same cost structure, mechanisms and

threats, First Tiers explore a smaller part of information space

It ensures a similar profitability when agents do not cheat, and a greater

profitability when they cheat

First Tiers

Final producersC

heati

ng

ag

en

ts in

RE

BT

Ch

eati

ng

ag

en

ts in

RE

BT

RNDREBT.0 REBT.5 REBT.1

100

130

160

190

220

250

280

310

0 100 200 300 400

100

130

160

190

220

250

280

310

0 100 200 300 400

vs.

Lucio Biggiero 2009

• How are tacit and explicit knowledge created, cumulated and transferred between organizations (firms and centers of research).

• What is the role of innovation and imitation?• What is the role of bounded rationality?• Etc.

A set of structural

and behavioral variables

K creation and transfer

Successo competitivo

?

The KNOWTIC Model

Spontaneous dynamics or policy interventions

Competitiveness at organizational,

inter-organizational,

and cluster level

Industrial cluster

Lucio Biggiero 2009

Some research hypothesesH1: La capacità di assorbimento permette di colmare il gap

conoscitivo esistente tra due distretti/cluster industriali.

H2: La capacità di assorbimento influenza le strategie di investimento in ricerca e sviluppo.

H3: La capacità di assorbimento permette di ottenere risultati vantaggiosi anche in presenza di costi di R&D elevati.

H4: Una distribuzione di capacità di assorbimento tra le ditte di un cluster che segue una funzione di potenza (80/20) produce più conoscenza di altre distribuzioni.

Lucio Biggiero 2009

The virtual experiments to test the first hypothesis

Per testare questa ipotesi vengono condotte 5 simulazioni facendo variare tre fattori:

  H1s1 H1s2 H1s3 H1s4 H1s5Capacità di assorbimento bassa bassa alta alta bassaConoscenza iniziale posseduta bassa alta bassa bassa altaImpatto conoscenza sulle spese in R&D basso basso basso alto alto

Quantità di conoscenza posseduta da due distretti industriali H1s2 e H1s3 nell’arco di 100 intervalli.

Quantità di conoscenza posseduta da due distretti industriali H1s4 e H1s5 nell’arco di 100 intervalli

Quantità di conoscenza posseduta da due distretti industriali H1s2 e H1s3 nell’arco di 100 intervalli.

L’ipotesi di ricerca risulta essere confermata. E’ possibile colmare il gap conoscitivo esistente tra due distretti/cluster industriali attraverso investimenti volti ad incrementare la capacità di assorbimento, la quale incrementa in maniera esponenziale i suoi benefici in presenza di investimenti in ricerca e sviluppo elevati.

Lucio Biggiero 2009

Se riflettiamo onestamente e attentamente, la maggior parte delle

cose che insegniamo e che non è stata ottenuta per via sperimentale (reale o

virtuale) ènel migliore dei casi

vera ma non si sa veramente perché

altrimenti falsa

Lucio Biggiero 2009

5) Some suggestions for future works

• Esistono delle norme individuali (micro-behavior) e semplici in grado di indurre fenomeni sociali (macro-behavior) di sviluppo sostenibile?

• Esistono delle norme individuali (micro-behavior) e complicate o forti in grado di indurre fenomeni sociali (macro-behavior) di sviluppo sostenibile?

Lucio Biggiero 2009

• Esistono delle norme individuali (micro-behavior) e semplici in grado di scoraggiare fenomeni sociali (macro-behavior) di sviluppo sostenibile?

• Esistono delle norme individuali (micro-behavior) e complicate o forti in grado di scoraggiare fenomeni sociali (macro-behavior) di sviluppo sostenibile?

Lucio Biggiero 2009

In ciascuno dei 4 casi precedenti, quali sono i fattori chiave?

Le soglie?

Le preferenze individuali?

Le asimmetrie informative?

La razionalità degli agenti?

Le loro interdipendenze?

Lucio Biggiero 2009

Che cosa cambia se si introduce un attore collettivo?Che caratteristiche deve avere per essere efficace?

Che ruolo gioca il grado di (de)centralizzazione decisionale?Che ruolo gioca il grado di (de)centralizzazione della produzione (o del consumo) di energia, materie prime, cibo, ecc.?

Lucio Biggiero 2009

Che ruolo gioca la differenziazione (economica, sociale, geografica, culturale,

ecc.) dei produttori e dei consumatori?

Esistono scale effects?

Does topology matter?

Che ruolo giocano opportunismo, attitudine cooperativa, ecc.?

Lucio Biggiero 2009

La simulazione ad agenti è lo strumento ideale e il più

appropriato per tutti gli studi di scenario e di policy