Directions of Emergence. Reputation and Social Norms Rosaria Conte LABSS/ISTC-CNR AISB, Aberdeen, UK, April 1- 4, 2008.

Directions of Emergence.Reputation and Social Norms

Rosaria Conte

LABSS/ISTC-CNR

AISB, Aberdeen, UK,

April 1- 4, 2008

Emergence

An effect is said to be emergent when it is generated by micro-level entities in interaction.

Apart from debates on • Properties (such as unintentional, unpredictable, unreducible; Kim, 1995), etc.• Orders of emergence: (1st-order and 2nd-order emergence: consciousness for

Dennett: is a second order emergence, in the sense that it emerges from the interactions of the parts of the mind and, through its emergence, changes how the system processes information.

• Etc.

it is generally perceived as an upward process What about way back?

• Existing theories of downward causation (such as Emmeche et al,, 2000) are affected by debate about reductionism and even by metaphysics (Abdoullaev, )

• A downward notion of 2nd-order emergence has been put forward (Gilbert), as implying recognition of emergent effect (see also, Goldspink and Kay, in ongoing symposium): emergent effect is represented by agents, thus contributing to its replication.

Need for General Theory of Downward Causation

Direct influence on behaviour (see Gilbert, 2002), new properties at lower level (e.g., stigma, exchange power,etc.).

2nd order emergence, as recognition of emergent effect, which contributes to replicate it (clustering in segregation model, Gilbert, 2002). Reputation is another example.

Immergence: effect cannot even emerge unless it Immerges into the mind of generating entities’ (Castelfranchi,

1998, Andrighetto et al., 2008) modifying representations and operating rules. Norms are one example.

Reputation

From Image to Reputation through Gossip Reputation is the

Emrgent effects (= reported on evaluation) of a

Social process (gossip) Starting from social

evaluation ( I = image) Twofold effect

On target: stigma On gossipers:

• social meta-belief about others’ evaluations

• Through multiple loops: (meta-)belief > gossip > retroacton > meta-belief/stigma, etc.

Why Bother?

For evolutionary theorists (Dunbar, 1998; Panchanathan, 2001), reputation allowed the

evolution of indirect reciprocity and the

enlargement of hominids’ settlements

Figura Interplay between informational and mateiral, direct and indirect reciprocity. Reproduced from Conte and Paolucci (2002).

Step 1: Direct material reciprocity

A B

give

Step 2: direct informational reciprocity

A C

tell

…A nB A

tell tell tell

Step 4: indirect informational reciprocity

give give

…A nB A

give

Step 3: Indirect material reciprocity

As a Meta-Belief…

No personal commitment of speaker about nested beliefs’ truth value.

No responsability about their credibility (“I am told that…”)

Implicit source of rumour Indefinite author of evaluation

Image = social evaluationReputation = meta-evaluation. This implies:

Rumours spread even when nobody believes them!

Simulation-based Exploration

What effect do such cognitive differences bear?

Thanks to REPAGE, a tool developed at LABSS (Sabater et al., 2005; see EU-funded eREP Project, http://erep.istc.cnr.it/ )

Simulations on multiagent stylized scenarios

REP-AGE

Memory includes Predicates from

Experience (contract fulfilments) Communication from others (I

and R) Organized in a network of

dependencies, specifying which predicates contribute to the values of others: each predicate has a set of

antecedents and a set of consequents.

With new inputs, thanks to Detectors, if an antecedent is created, removed, or its value changes, predicate value is recalculated and change notified to its consequents.

REPAGE runs on a JADE-X platform.

Simulations run (Paolucci et al., 2007; Quattrociocchi et al., 2008) in simplified markets, to explore trade-off of informational cooperation:

Communication is necessary to find good sellers. But agents have an incentive to cheat.

Hence: Fixed number of sellers and buyers (respectively to 100 and

15), Goods are represented by a 1-100 valued utility factor Variable quality sellers with finite stocks, which, when exhausted, are

replenished automatically with random quality. Buyers

purchase, Ask for info from one another (which is the best, which is the worst) Answer by providing

false/truthful info (info cheating rate) Tested (I) or untested ® information

Simulations with REPAGE

Experimental Conditions

• L1: market with only Image• L2: same market with Image + Reputation Explore respective performance, considering that

in L1 either Tested image spreads, or Retaliation (when false image is transmitted),

In L2, less retaliation is expected and more, although untested, information.

The two curves present a different cyclic behaviour:

in L1 (blue) agents find more good sellers than in L2 (red).

Peaks of each wave is interpreted as exhaustion of stocks: once a good seller is discovered, buyers start to buy from this one until extinction of stock.

Minimum value for each wave is interpreted as a slow process of discovery.

Findings. Find Out Good Sellers

Average QualityAverage products quality in 100 turns. L1 (blue, I only) L2 (red, I + R). Both achieve optimal quality, with faster L2 convergence.

How is it possible?• Information spreadsmore in L2:

• Agents find less good sellers • do not exhaust them• try more: information circulates more (and more widely)

• but what info quality?

Uncertainty Vs Quality: L1

Uncertainty (= “I DONT KNOW” answers) grows with quality

Uncertainty Vs Quality: L2

In L2 (I + R), opposite correlation:, uncertainty decreases with growing quality

Evolution of uncertainty (I Don't Know) in 100 turns with only image circulating: values remain constantly high.

Evolution of Uncertainty: L1

Evolution of Uncertainty: L2

Preliminary Conclusions

Although both achieve good quality But with reputation

Uncertainty decreases: information does not get lost No exhaustion of resources

What is the use of reduced uncertainty, if quality is the same?

Results indicate three directions for further exploration Reputation might favour and be compatible with larger

networks (effect to be checked with open networks) Information can be transmitted to future generations

(effect to be checked with evolution of the market, spin-off, etc,).

What about not only scarce but also finite resources? One might think that reputation is more robust than image with non-self replenishing resources.

Norms

Two Current Views

Conventions (mainly bottom-up) Legal norms (mainly top-down) Open questions

As to conventions: • What about social norms?• Why are they enforced?• What about mandatory social norms?

As to legal norms• How do they evolve?• How do agents find them out?

As to both• What about a unifying view?

2-way Dynamics of Social Norms (EMIL project: http://emil.istc.cnr.it/ )

Norm: a behaviour that spreads thanks to the spreading of normative beliefs and commands

Normative belief: a belief that a given action , in a given context, for a given set of agents, is forbidden, obligatory, permitted, etc.

Normative command: a command based upon a normative belief (more precisely,a command that wants to be adopted via the formation of a normative belief).

The Input

Each input is presented as an ordered vector consisting of four elements:

Source (x); Modal (M) through

which the message is presented: assertions (A), behaviours (B), requests (R), deontics (D), evaluations (V), sanctions (S);

Observer (y); Action transmitted (α).

N-Recognition Module

B, R, A

> vc

D, V < vc

N-bel

Input

N-Board

E

Board ofAuth.

YN

Why Bother? Simulations of norm-recognizers against social conformers

in different populations (Campennì et al., 2008a, papers submitted to WCSS; 2008b, submitted to NORMAS)

The model: Multi scenario world

• Four different multi-action scenarios (social settings)• With one common + two scenario-specific actions (total nine

actions). Agents

• move from one scenario to the next• are endowed with

• Personal agendas • Individual fixed time of permanence in each scenario• Two populations

• Social conformers: follow actions most frequently done in observation window (parameter).

• Norm recognizers take input from others, form beliefs and act based on those.

Preliminary Findings Each colour represents one action Social conformers:

No difference within ticks Strong difference

• Among ticks (no belief)• Among scenarios (no memory)• More frequent action (dark blue) is

distributed throughout the simulation: nothing emerges!

Norm recognizers: Fuzzier

• Rows (autonomy)• Columns (beliefs)

After 60th ticks, one action common to all scenarios: something emerges…

What is it? Lets look into agents beliefs…

Immergence

At the 30th tick a normative belief starts to spread as well

Immergence is earlier: it takes time for effect to emerge (loops).

What has happened in the meantime? Other normative beliefs were formed,

although earlier is more frequent If same-norm agents get separated

(genetic or cultural drift): norm innovation! (equally frequent norms might emerge in different subpopulations).

If they then get re-united, which norm is going to invade population?

Question for future studies :-)

Final Remarks Macrosocial regularities emerge and modify the generating

machines. Different types and degrees of top-down influence:

agents recognize emerged effects sometimes effects don’t emerge unless they immerge.

Hence, we need to understand this process to Understand agents Understand different patterns of macrosocial regularities:

• With reputation, observable marosocial effects of reduced uncertainty might include larger networks, higher stability, more robustness.

• With social norms, observable macrosocial effects of normative beliefs • actually include effective convergence across scenarios,

• Potentially, norm-innovation

References Andrighetto, G., Conte, R.,Turrini, P., Paolucci, M. (2007). Emergence

In the Loop: Simulating the two way dynamics of norm innovation. In Proceedings of the Dagstuhl Seminar on Normative Multi-agent Systems, 18-23 March 2007, Dagstuhl, Germany.

Andrighetto, G., Campennì, M, Conte, R., Paolucci, M. (2007). On the Immergence of Norms: a Normative Agent Architecture. In Proceedings of AAAI Symposium, Social and Organizational Aspects of Intelligence, Washington DC.

Conte, R., Andrighetto, G., Campennì, M, Paolucci, M. (2007). Emergent and Immergent Effects in Complex Social Systems. In Proceedings of AAAI Symposium, Social and Organizational Aspects of Intelligence, Washington DC.

Andrighetto, G., Campennì, M, Conte, R., Cecconi, F. (2008). Conformity in Multiple Contexts: Imitation Vs Norm Recognition, The second World Congress on Social Simulation (WCSS-08), George Mason University, Fairfax - July 14-17, 2008. Submitted.

Andrighetto, G., Campennì, M, Conte, R., Cecconi, F. (2008). How Agents Find out Norms: A Simulation Based Model of Norm Innovation, 3rd International Workshop on Normative Multiagent Systems (NorMAS 2008), Luxembourg, 15-18 July, 2008. Submitted.

Andrighetto, G.; Campennì, M.; Conte, R. (2007). EMIL-M: MODELS OF NORMS EMERGENCE, NORMS IMMERGENCE AND THE 2-WAY DYNAMIC, Technical Report, 00507, LABSS-ISTC/CNR.