1 kimas 2003dr. k. kleinmann an infrastructure for adaptive control of multi-agent systems dr. karl...
TRANSCRIPT
1KIMAS 2003Dr. K. Kleinmann
An Infrastructure for Adaptive Control of Multi-Agent Systems
Dr. Karl Kleinmann, Richard Lazarus, Ray Tomlinson
KIMAS, October 1, 2003
2KIMAS 2003Dr. K. Kleinmann
Control of DMAS: Problem Description
• Characteristics of Distributed Multi-Agent Systems (DMAS) reason that formal methods of control theory are rarely applied in software engineering for agent systems
• “What makes DMAS a hard (complex) control problem?”– Dynamic system boundaries, interactions and communication
paths– System size, number of internal states, degrees of freedom
(flexibility)– Strong couplings between system states (shared resources)
• “What makes DMAS a unique control problem?” – Shape of cost functions and performance criteria
• Changes of control inputs at (almost) no cost (nonlinear impact)
– Explicit model available, accurate description of behavior (code)• System as its own simulator
– Extensive experimentation “for free” (automated testing)• Control approaches and parameter variations by “trial and
error”
3KIMAS 2003Dr. K. Kleinmann
Cougaar Project Background and Control Objectives
• Paper presents the Control Infrastructure of the Cougaar Distributed Agent System
• Cougaar is an Open Source Agent Infrastructure developed under the DARPA Programs– ALP (1996-2001): Military Logistics Planning– UltraLog (2001-2004): Survivable Logistics Planning and Execution
• Primary System Function– Logistics Plan
• Robustness Function– Maintain Processing Infrastructure
despite Loss of Resources• Security Function
– Maintain System Integrity despite Information Attacks
MilitaryLogisticsOperation
UltraLogDMAS
Logistics Requirements
Logistic Actions (Plan)
SW Failures HW Failures
4KIMAS 2003Dr. K. Kleinmann
Cougaar Architecture Overview
• 100% Java Architecture for building large DMAS
• Proven Scalability– Prototype with 500 distinct
agents distributed over 5LAN network of >100 machines
• Two-Level interaction model– Intra-agent blackboard for
tightly-coupled interactions– Inter-agent message passing
for scalable loosely-coupled interactions
• Distributed object management– Prototype/delegation data
model– Capabilities-based
representations• Two-Dimensional containment
model– Components can be both
containers and plugins
PLATFORM OS
JAVA VMJAVA VM
COUGAAR NODE
Agent Binder
Agent Binder
PLATFORM SERVICES
COUGAAR NODE SERVICES
Agent Binder
AGENT.AGENTFRAMEWORKSVCS.
PluginBinder
PluginBinder
PluginBinder
PluginBinder
PlugInPlugIn
COUGAAR NODE
Agent Binder
COUGAAR NODE SERVICES
Agent BinderAgent Binder
AGENT.AGENTFRAMEWORKSVCS.
PluginBinder
PluginBinder
PluginBinder
PluginBinder
Agent Binder
AGENT.AGENTFRAMEWORKSVCS.
PluginBinder
PluginBinder
PluginBinder
PluginBinder
PlugInPlugIn
PlugInPlugIn
PlugIn
PlugIn
TRANSCOM
1BDE
2BDE
5KIMAS 2003Dr. K. Kleinmann
Control Levels designed to achieve Control Objectives
• Agent Infrastructure-Level Control– Parameters of Components within an
Agent– “Local Agent Autonomy” (e.g., Message
Compression; Status Report Rate)
MilitaryLogisticsOperation
UltraLogDMAS
Logistics Requirements
Logistic Actions (Plan)
SW Failures HW Failures
Agent Components
AgentController
TRANSCOM
1BDE2BDE
UltraLog DMAS
• Application-Level Control– Complex Actions or Sequences of Actions
composed of Control Primitives– “Specific Defenses against Stresses”
(e.g., Load Balancing; Agent Restart)
6KIMAS 2003Dr. K. Kleinmann
The Cougaar Control Infrastructure
ProcessingComponents
(Plugins)
Sensors
Publish real-time sensor conditions, e.g., Load, Resource AvailabilityTHREATCON
Sets Operating Modes for Components based on plays in Playbook and current Sensor Conditions
Read Playbook
Constrain Playbook based on OperatingMode policy direction
Operating mode policy manager reads Operating mode policies (relayed from other agents) from blackboard
Publish OperatingModes and TechSpecsGet Condition
by name
Get OperatingModeby name
Blackboard
RelayLP
Other agents
Inter-AgentOperatingMode Policies
Operating ModePolicy Manager
Adaptivity Engine
OperatingModeService
ConditionService
Playbook
PlaybookConstrain Service
PlaybookRead
Service
Playbook Manager
ProcessingComponents
(Plugins)
ProcessingComponents
(Plugins)
SensorsSensors
Publish real-time sensor conditions, e.g., Load, Resource AvailabilityTHREATCON
Sets Operating Modes for Components based on plays in Playbook and current Sensor Conditions
Read Playbook
Constrain Playbook based on OperatingMode policy direction
Operating mode policy manager reads Operating mode policies (relayed from other agents) from blackboard
Publish OperatingModes and TechSpecsGet Condition
by name
Get OperatingModeby name
Blackboard
RelayLP
Blackboard
RelayLPRelayLP
Other agents
Inter-AgentOperatingMode Policies
Operating ModePolicy Manager
Adaptivity Engine
OperatingModeService
ConditionService
Playbook
PlaybookConstrain Service
PlaybookRead
Service
Playbook Manager
Playbook
PlaybookConstrain Service
PlaybookRead
Service
Playbook Manager
7KIMAS 2003Dr. K. Kleinmann
Example (Part of the Open Source Code Base)
• 2 Agents– (Consumer); Provider
• 2 Sensor Inputs (Conditions)– Task Rate; Avail. CPU
• 1 Control Input (Operating Mode)– Task Allocator Plugin Mode
(Tradeoff Accuracy vs Speed)
Task AllocatorPlugin
AdaptivityEngine
Task Rate CPU
Allocation Alg.
Prov
ider
Age
nt
Cons
umer
Age
nt
Sensor Values (Conditions) over time
0.0
2.0
4.0
6.0
8.0
10.0
12.0
14.0
16.0
18.0
20.0
1 5 9 13 17 21 25 29 33 37 41
Task Rate(Tasks/sec)
Available CPU (%/10)
Control Input (Operating Mode) over time
0
100
200
300
400
500
600
1 5 9 13 17 21 25 29 33 37 41
Iterations (cycl/sec)
8KIMAS 2003Dr. K. Kleinmann
Interpretation of the Approach
• Infrastructure allows both feedforward and feedback control, depending on selection of Conditions and Operating Modes
• Heuristic parametrization of Plays makes Adaptivity Engine typically nonlinear controller (-> fuzzy control)
• If Plays in Adaptivity Engine are modified according to TechSpecs or constrained by Policies, Control System becomes truly “Adaptive”
• Rule Matrix in Example:
AE Playbook (Control Algorithm)
0
200
400
600
1 51 101 151 201
Condition (Task Rate / CPU )
Op
erat
ing
Mo
de
(Allo
cati
on
It
erat
ion
s)Avail. CPU Allocations
high max med
low med min
low high Task Rate
9KIMAS 2003Dr. K. Kleinmann
Conclusions
• Presented Agent-level Control Infrastructure for the Open Source Cougaar Architecture– Generic Approach allows to address Control
Objectives of a DMAS Survivability Application– Connects Software Engineering with Control
Theory Models in Order to leverage Control Theory Methodologies
• Future Research Issues – Components publishing TechSpecs– Deconfliction of Application-level Control
Strategies
10KIMAS 2003Dr. K. Kleinmann
For more information …
• BBN Technologies: http://www.bbn.com• Cougaar: http://www.cougaar.org• UltraLog: http://www.ultralog.net• Other Cougaar-related KIMAS’03 papers:
– “Multi-Tier Communication Abstractions for Distributed Multi-Agent Systems”, M. Thome, et al
– “Multi-resolutional Knowledge Representation for Logistics Systems using Prototypes,Properties and Behaviors”, J. Berliner, et al