Emergence of Norms through Social Learning

IJCAI’07

Emergence of Norms through Social Learning

Partha Mukherjee, Sandip Sen and Stéphane AiriauMathematical and Computer Sciences DepartmentUniversity of Tulsa, Oklahoma, USA

ALAg-07

Introduction

Norm: “a convention as an equilibrium that everyone expects in interactions that have more than one equilibrium” [Young, 1996]

Use a population of learning agents to simulate a population that faces a

problem modeled by a game and study the emergence of norms

Use a population of learning agents to simulate a population that faces a

problem modeled by a game and study the emergence of norms

ALAg-07

Example of a norm: picking the side of the road

Agents need to decide on one of several equally desirable alternatives.

This game can be extended to m actions

4

4

-1

-1

-1

-1

4

4

L

R

R L

ALAg-07

Previous Work

Previous work on learning norms assume observation of other interactions between agents.

How norms will emerge if all interactions were private?

Social Learning (IJCAI-07): agents play a bimatrix game, at each interaction, an agent plays against another agent, taken at random, in the population

Empirical study: Study effect of population size, number of actions available, effect of learning algorithms, presence of non-learning agents, multiple relatively isolated populations

ALAg-07

Social Learning

Population of N learning agentsA 2-player, k-action game M M is common knowledgeEach agent has a learning algorithm

(fixed, intrinsic) to play M as a row or a column player

Repeatedly, agents play the game M against an unknown, random opponent.

ALAg-07

Protocol of play

For each iteration, for each agentPick randomly one agent in its

neighborhoodFor each pair, one agent is randomly

considered row, the other column playerEach agent pick an action, and can observe

only the action of the other agent constituting the pair

Each agent gets the reward accordingly, and updates its learning mechanism

ALAg-07

Interactions are limited to neighboring agents

ALAg-07

Effect of neighboring size

ALAg-07

Learning Dynamics

D=1

D=15

It 145 It 355 It 480

Driving on the left Driving on the right

ALAg-07

Influence of non-learners

Non-learners use identical strategiesD=5

ALAg-07

Influence of non-learnersUsing different strategies

Driving on the left Driving on the right

D=1

D=15

It 45 It 535 It 905

ALAg-07

Conclusion

Bottom up process for the emergence of social norms

Depends only on private expertise Agents can learn and sustain useful social

norms Agent population with smaller neighborhoods

converge faster to a norm