On the Modeling, Refinement and Integration of Decentralized Agent Coordination

Jan Sudeikat1,2 [email protected] Renz1 [email protected] of Applied Sciences Hamburg - Multimedia Systems Laboratory 2University Hamburg - Distributed Systems and Information Systems

Distributed Systemsand Information Systems

On the Modeling, Refinement and Integration

of Decentralized Agent Coordination– A Case Study on

Dissemination Processes in Networks

2009-03-25

International Workshop on Self-Organizing Architectures (SOAR 09)

Cambridge, UK

mailto:[email protected]

mailto:[email protected]

On the modeling, refinement and integrationof decentralized agent coordination


Distributed Systems Architectures Challenge:

Building adaptive applications that are scalable, robust, …

Architectural Choices: Managed Hierarchical

Decentral

Local adaptive entities: software agents Problematic: effective coordination

Here:Utilization of

Self-Organizing Processes

Scalability,Robustness, …

Managing Entity

Pyramid of Managing

Entities



Self-Organization as a (Software) Design Principle

Self-Organization: physical, biological and social phenomena, global structures arise from the local interactions

of autonomous individuals (e.g. particles, cells, agents, ...) Structures are:

Adapted to changing environments Maintained while being subject to perturbations

Attractive for software architects: Decentralized coordination strategies / mechanisms

No single point of failure Conceive application dynamics resemble phenomena Blending of functionality and coordination aspects (Reuse, Redesign)

Requirement: Systematic conception / integration Declarative configuration of agent coordination Enactment architecture

(Sudeikat & Renz 2008, 2009)



Proposal: Programming Model for Self-Organization

Self-organizing processes result from coupled feedbacks between system elements

Context dependent amplification / damping

of element activities

Systemic Modeling Approach System Science concepts characterize MAS operation

System Variables: # behavior exhibitions (roles, groups, …)

Causal Relationships: rates of variable changes Feedback-Networks

Toolset: Configuration Language Enactment Architecture

…

+(+)

reinforcing

+

++

+(-)

+

++

-/ balancing



Coordination Enactment Architecture

Layered Approach Application Coordination

Coordination Media Interaction techniques

Agent-Modules Execution Infrastructure

Coordination-Endpoint: Agent-modules Interface Coordination Media

Publish / Subscribe mechanism Automating coordination-activities

1: Agent observation / modification 2: Controlled by coordination model 3: Publication of agent adjustments

Externalized Coordination

Model



Coordination-Endpoint: Agent State Interpreter

Observe agent execution Behavior-Classification Behavior-Change Publication

Coordination Information Interpreter

Reception via CM. Adjustment of agent-behavior

Local Adaptivity: Declarative: Conditions / Invariants Adaptivity Component: (optional) Procedural Implementation of

Classification of Observations Adaptations of Agent state

Coordination Medium Publish / Subscribe Interface

Coordination Enactment Architecture

Realizing self-organizing processes: Information FlowsLocal Element Adaptivity



Methodic Conception of SO-Processes Integration of Coordination Development in AOSE

AOSE: Tools / techniques for agent development Plan for concerted phenomena

Systematic refinement procedure Describing System Behavior1. Identify Problem Dynamic

Structures Attractors coupled feedback loops

2. Propose Solution Dynamic Opposing / Corrective Structure

3. Refinement operations Map Coordination model to Agent models



Case Study I: Convention Emergence Decentralized agreement problem in MAS

Communication of local settings Agents adjust accordingly

Embedding an externalized Coordination Model Generic agent activity

Coordination Model: Observation of activities Communication of configurations Adjustment Policy: majority rule

+/- feedback loop Coordination Medium: Overlay-Network Topology

Convergence



Case Study I: Convention Emergence Sample Simulation Run:

Random Initialization Value Convergence

Random agent activation Communication: Coordination Medium

Impact of Network-Topology: Random Graph Power law Graph: Comparable convergence times

Less communicative overhead in power law distributed graphs



Case Study II: Patching Dynamics Exemplify refinement process: Problem description correcting coordination process

Problem: Spreading of “infections”

in agent population Agent exhibit two Roles:

Susceptible Infectious

Balancing vs. reinforcing Feedback Goal-Seeking

Possible Solution Dynamic: Additional Balancing Feedback Limit Susceptible and Infectious agents



Case Study II: Patching Dynamics Refined Solution Dynamic

Executable! Adaptivity Component

Functionality Behavior Classification

Information Flow

Sample Simulation Run One random infection Fixed infection rate

Epidemic Recovery of initial infection

starts recovering process

unsusceptibleinfected



Conclusions I

Embedding of self-organizing processes in MAS

Architectural Aspect: Proposal:

Reference Architecture Declarative language support

Supplement Coordination Encapsulation of:

Adaptation logic Information Flow / Interaction Technique

Methodic Aspect: Equip self-organizing process to

correct / oppose problematic dynamics



Conclusions II

“… how their contribution connects the self‐adaptiveperspective with the self‐organizing perspective”

(System) Self-Adaptivity by concerted entity adaptivity Adaptive Software System:

Establishment of closed feedback loop, e.g. MAPE, … Here:

Collective adjustments of individual elements Closed feedback is distributed among system elements

System coordination model Sets of feedback loops



End

Thank you for your Attention!

Questions / Suggestions are welcome!



Case Study I: Convention Emergence Sample Simulation Run:

Random Initialization Value Convergence

Random agent activation Communication: Coordination Medium

Impact of Network-Topology: Random Graph Power law Graph: Comparable convergence times

Less communicative overhead in power law distributed graphs



Encapsulating Adaptivity / Interaction

Foundational elements of a self-organizing processes Information Flows Local Element

Adaptivity

Coordination Media: Information exchange techniques

Tuplespace, spatial environments,… Here, Overlay-Network

Topology constraints communication

Coordination Endpoints: Local adpatation knowledge Automation of coordination-related activities



Coordination Pattern



Systemic Software Modeling



Modeling Notation



Systemic Modeling Causal relations of system variables

Describe Entity behaviors Anticipation of the Qualitative System Dynamics

Manual inspection and/ or simulation

A Hypothetical System: Producers Products Products Storage Storage Production

Exemplifying Systemic Modeling of MAS

Balancing Feedback

Practical development: After a suitable causal structure has been found:How to implement ?



MASDynamics: Declaration of Agent Behavior Interdependencies Systemic system model:

Nodes System Variables # of role occupations # of groups …

Interdependencies: Links Direct:

e.g. service invocations, … Mediated:

using environment models, e.g. pheromones, tuple spaces, … Description levels:

Application independent Alignment with agent implementation:

Node Types LinkTypes

Nodes: Referencing reasoning events that indicate behavior adjustments, E.g. goal adoptions, plan activations, …

Links: Configuring interaction techniques E.g. environment models, …



Coordination Strategies Systemic Modeling of macroscopic dynamics

Compensating Amplifying Selective




Compensating:




Amplifying:




Selective:



Decentralized Coordination Mechanisms

Information Exchange techniques Classification:



Expressing Coordination Dynamics Structural Properties of SO-Systems

Positive Feedback Amplification of appropriate entity activities

Negative Feedback Damping inappropriate entity activities

... Dynamic Viewpoint on application development:

Consider dyn. properties at design-time Design the causes of self-organization

MAS specific modelling level: Agent-based design concepts:

Roles: Abstraction of agent behaviours Groups: sets of individuals that

share common characteristics (e.g.: collective goals)

System State: # of behaviour occupations



Case Study: Decentral Web-Service Management

Agent-based Web-Service Management Architecture

Balance service workloads Management Agents:

(J2EE) Service-Endpoint Broker Agents

Registries: Service-Endpoints

Prototype Implementation: Jadex Agent Platform

Cognitive agent model Beliefs, Goals, Plans, Internal Events, …

SUN Appserver Management Extensions (AMX) Server-Management Interface

Conceptual Architecture

http://jadex.informatik.uni-hamburg.de/bin/view/About/Overview

https://glassfish.dev.java.net/javaee5/amx/



Case Study: Decentralized Web-Service Management

A Functional, but un-coordinated Implementation

Manual management of is enabled Tropos Modeling Notation Dependencies of agent types

Client Service Endpoint Client Broker Broker Service Endpoint Broker Client

Systemic Description of the Causal Application structure

Accumulative system variables

Complementing the causalities Establish a negative feedback loop

Agent state definitions Establishment of interdependencies

Tropos Design Notation



Case Study: Decentralized Web-Service Management

Embedding Coordination: Strategy Definition:

Variable / Link Declarations Strategy alignment / integration

Referencing agent models

Configuring interaction technique

Validation: Provoking the manifestation

of the feedback loop Responsive regime Sudden demand for

specific service type

Event Publications

Event Perceptions

Middleware Configuration



Case Study: Behavioral Analysis by Applying Stochastic Process Algebra

Stochastic Process Algebra: Behavioral modeling System of interacting processes Coupled by synchronized activities

Validation of qualitative dynamic: Provoking the effects of

the feedback loop Responsive regime Initial Conf.:

Allocation of service 1 Input:

High demand of service 2

Balance of allocations



Mesoscopic Modeling Available formalisms:

Macroscopic System System Sciences Mathematics, …

Microscopic System Local entity (inter-)actions State Machines, Process Algebra, …

Transition: Simulation / Iteration of microscopic models

Proposal: (Renz & Sudeikat, 2005, 2006) Intermediate description levels: Mesoscopic agent states Classification of agent behaviors

Relevance of agent activities with respect to the Macroscopic System

Behavior Abstraction of the microsopic agent activities

Mesoscopic agent states: Not microscopic:

Coarse grained agent activities Not macroscopic:

Exhibits short time fluctuations



Applying Mesoscopic Modeling

Top-Down: E.g.: MASDynamics

Transfer of System Dynamics concepts

Graph-based modeling

Bottom-up: E.g.: Stochastic Situational Calculus

Extension of the Sit. Calculus

Two orthogonal approaches: Different modeling directions Enabling iterative development:

Explain rising phenomena Tune rising phenomena

modeling macroscopic dynamics refinement to intermediate scales

coarse-graining element dynamics inferring collective system properties



Top-Down: Systemic MAS Modeling MAS abstraction by:

Agent-based design concepts: Roles: Abstraction of agent behaviours Groups: sets of individuals that

share common characteristics (e.g.: collective goals)

Global MAS State: # of behaviour occupations

Graph Definition: Nodes: System Variables

# of role occupations # of organizational groups size of organizational groups quantification of environment elements ( #, size, etc. )

Links: Causal relations Environment mediated Direct agent interactions

Modelling the causes of Self-organization: Feedback Loop Structures

MAS Design



Top-Down: Systemic MAS Modelling Allows for model refinement

Attachment: add detail Link: detail link dynamics Variable: detail variable intern

dynamics

Example: Ant-based path finding

(-)

(+)



Self-Organization vs. Emergence Methodological view

On the Modeling, Refinement and Integration of Decentralized Agent Coordination

Documents

adjustment of agent

coordination modelon

coordination model3

coordination aspects

software architects

software agents problematic

dissemination processes

adaptive applications