Artificial Agent Society Simulations in an Encounterbased Normative Action Environment Hrevren Kili

1
Artificial Agent Society Simulationsin an
Encounter-based Normative Action Environment
Hürevren KiliçComputer Engineering
DepartmentAtilim University, Ankara,
TURKEYhurevren_at_atilim.edu.tr
2
CONTENT

• Introduction
• Agent and System
• Hypothesis and Developed Measures
• Simulation Setup and Results
• Discussions and Conclusions

3
INTRODUCTION

• A design consideration in MultiAgent Systems
  (MAS)
• Behavioral variance (?)
• Respect to individual autonomy
• Norm socially enforced rules activated during
  encounters, useful in behavioral variance
  reduction conflict resolution.
• Encounter and Norm based Action Environment.

System Level Efficiency
4
INTRODUCTION

• Question Is there any dependency between agents
  socialness and societys behavior predictability
  under different autonomy degree settings ?
• Approach Artificial Agent Society Simulations.
• MAS setting
• Cooperative normative.
• No spatial awareness.

5
INTRODUCTION

• Problems
• What will be the generic problem to be solved by
  agents ?
• How to represent single agent ?
• How norms will spread among agents be
  internalized by an agent ?

6
INTRODUCTION

• Solutions
• Single-source Single-destination Shortest Path
  (SSSP) Problem.
• Self and Group Models.
• Through encounters. Development of measures for
  norm spreading and internalizations

7
AGENT and SYSTEM

• Definition Single-source Single-destination
  Shortest Path (SSSP) problem is defined by three
  components G(V, E ,w) s and d where G(V, E, w)
  is a weighted directed graph composed of V set
  of vertices, E set of edges and w weight
  function E ? R s is the source node and d is
  the destination nodes of the problem such that s,
  d V.
• SSSP polynomial-time solvable.
• Each agent assigned SSSP and different tasks of
  reaching from some s to some d.

8
AGENT and SYSTEM

• Basic characteristics of MAS setup are satisfied
• Partial observability of the environment.
• Agents cannot solve their SSSP in the best way by
  themselves.
• No centralized global control over the system.
• Asynchronous communication through random
  encounters.
• No central shared data repository but data
  acquisition through encounters.

9
AGENT and SYSTEM

• Agents are rational.
• 0 Agents Autonomy (a) 1. (Fixed!)
• Self-model SSSP Self knowledge about the
  environment assigned task.
• Group-model SSSP Agents knowledge about the
  society.
• Initializations
• Self-model ? Group-model
• The same default Group-model for all agents.

10
AGENT and SYSTEM
Action Decision
Self Model

Updated self model
Action
Self model
Updated self model
Upper bound for of encounters
Model Merge Update
External Param.
Autonomy value
Group Model
Accumulated self models
Group Model
Group Model Revision
Revision period value
Updated group model
Other agents self models
Single agent design
11
AGENT and SYSTEM

• Self-model update
• if both edges exist
• if only self models edge exist
• if only group models edge exist
• None if none of the
  edges exist
• where a is the autonomy value.
• The source and destination nodes of the model
  Snew are the same as of Scurrent.

12
AGENT and SYSTEM

• Group-model update
• Group model update period (q) - external
  parameter.
• Group model is updated by using other agents
  memorized self- models by the end of every q
  encounters.
• No-edge instance is represented by 0.
• where parameter q N is related with the
  agents degree of socialness. Socialness is
  defined as (1/q).

13
AGENT and SYSTEM

• Lower q value is an indicator of high socialness
  degree or vice versa.
• The number of nodes of self and group models are
  equal.
• Total number of encounters in a simulation is
  controlled by a parameter.

14
HYPOTHESIS and MEASURES

• Lower degree of agent socialness results in
  lower behavior predictability of society.
• An average agents behavior variance can be
  calculated by norm internalization measure
• Norm_Int
• where m is the number of agents. is the
  number of nodes of the self and group model
  graphs. and are the edges
  connecting the ith node to the jth node of the
  self and group models, respectively.

15
HYPOTHESIS and MEASURES

• Another measure is to look at the differences
  between the edges of the group models accross the
  agents as below
• Norm_Spr
• where m is the number of agents. is the
  number of nodes of the self and group model
  graphs. is the mean group model. An edge
  (i, j) of group mean model is calculated by

16
HYPOTHESIS and MEASURES

• High variance of behavior (i.e. low behavior
  predictability) is recognized by
• Low norm internalizing factor (i.e. the agents
  have a higher difference between their self and
  group models),
• High norm spreading factor (i.e. agents are not
  close to each other with respect to their vision
  on the norms of the group).

17
SIMULATION SETUP

• Initialization Parameters
• Number of agents (set to 75)
• Number of nodes per model (set to 75)
• Max edge value (0 to 20)
• Bounds for the number of outgoing edges per node
  (2 to 10)
• Number of encounters (set to 1000) equilibrium
  is reached.
• Each simulation setup described by
• Autonomy values (0.0, 0.3, 0.6, 1.0)
• Socialness values (1/1, 1/5, 1/10)
• Independent initial model task assignments (5
  different)
• 4x3x5 60 independent setups.

18
SIMULATION RESULTS

• Using Norm Internalizing measure
• Hypothesis holds only for a1.0

• Using Norm Spreading measure
• Hypothesis holds only for a0.3 and 0.6

Existence of some degree of autonomy
together with lower degree of socialness ? ? ?
Lower behavior predictability of
individuals hence of society

19
DISCUSSIONS and CONCLUSION

• Problems
• Huge initialization space formed by of agents,
  of nodes per model, of edges, edge values,
  agent autonomy and socialness values
• Some hidden system behavior may not be revealed !
  
• Quality of random number generator.
• Prevention of occurences of some encounters !
• Repetitive (cyclic) encounter occurences limiting
  the investigation of whole encounter space !
• Very simple representations for autonomy and
  socialness.
• Better self and group model representations and
  update formulation.

20
DISCUSSIONS and CONCLUSION

• Description of potential basic components of an
  infrastructure for real world social modeling.
• Realization of infrastructure hypothesis
  testing in an artificial agent society setup.
• Future work
• Addition of spatial-awareness component.
• Use of the developed infrastructure for social
  modeling of real life entities (data required).
• Example problem Modeling and simulation of
  human wayfinding behavior in partially known
  environments.

21

• ! Thanks !
• ? Questions ?