View
214
Download
0
Tags:
Embed Size (px)
Citation preview
ADVIS '04 1
Artificial Life: How can it impact engineering practices of the
future? Cihan H. DagliCihan H. Dagli
Smart Engineering Systems LaboratorySmart Engineering Systems Laboratory
Engineering Management DepartmentEngineering Management Department
University of Missouri - RollaUniversity of Missouri - Rolla
Rolla, MO 65409 - 0370Rolla, MO 65409 - 0370http://www.umr.edu/~daglihttp://www.umr.edu/~dagli
ADVIS '04 2
Presentation Outline
Engineering Systems of the FutureEngineering Systems of the Future What is Artificial Life?What is Artificial Life? Artificial Life in EngineeringArtificial Life in Engineering Concluding RemarksConcluding Remarks
ADVIS '04 3
Recent Market Changes
Total GlobalizationTotal Globalization Increasing Production PaceIncreasing Production Pace Decreasing Production Cycle TimesDecreasing Production Cycle Times Migration From Mass Production to Mass Migration From Mass Production to Mass
CustomizationCustomization
ADVIS '04 4
Engineering Systems of the Future
Immediate Respond to Market ChangesImmediate Respond to Market Changes More Sensitive to Customer NeedsMore Sensitive to Customer Needs Migration from Central to Distributed Migration from Central to Distributed
ControlControl Autonomous and Cooperating Production Autonomous and Cooperating Production
Units Units
ADVIS '04 5
Smart Systems
The term “smart” indicates physical The term “smart” indicates physical systems that can interact with their systems that can interact with their environment and adapt to changes environment and adapt to changes through self-awareness and perceived through self-awareness and perceived models of the world, based on quantitative models of the world, based on quantitative and qualitative information.and qualitative information.
ADVIS '04 11
“Trajectories” of Research into Distributed Systems
System Behavior & Analysis
System Behavior & Analysis
System Design
System Design
Swarm Swarm Intelligence & Intelligence &
Synthetic Synthetic EcosystemsEcosystems
Artificial Artificial LifeLife
Multi-Multi-agent agent
SystemsSystems
Distributed Distributed Artificial Artificial
IntelligencIntelligencee
Population Population Biology& Biology&
Ecological Ecological ModelingModeling
ADVIS '04 12
What is Artificial Life?
A Perspective:A Perspective: It is a way of imitating Nature in order to solve It is a way of imitating Nature in order to solve
engineering problems. engineering problems. It includes simulation and emulation of living It includes simulation and emulation of living
systems like plants or animals. systems like plants or animals. It tries to achieve a new understanding of It tries to achieve a new understanding of
living systems, and of what is life.living systems, and of what is life.
http://kal-el.ugr.es/pitis.html
ADVIS '04 13
A Definition:A Definition:Artificial life is a field of study devoted to Artificial life is a field of study devoted to understanding life by attempting to abstract understanding life by attempting to abstract the fundamental dynamical principals the fundamental dynamical principals underlying biological phenomena, and underlying biological phenomena, and recreating these dynamics in other physical recreating these dynamics in other physical media – such as computers – making them media – such as computers – making them accessible to new kinds of experimental accessible to new kinds of experimental manipulation and testing.manipulation and testing.(by Christopher G. Langton, from the preface to the Proceedings of the Workshop on Artificial Life,
February 1990, Santa Fe, New Mexico)
What is Artificial Life?
ADVIS '04 14
Adaptive Autonomous Agents
Agent:Agent: A system that tries to fulfill a A system that tries to fulfill a set of goals in a complex, dynamic set of goals in a complex, dynamic environment.environment.
Environment:Environment: It can sense the environment through It can sense the environment through its sensors and act upon the its sensors and act upon the environment using its actuators.environment using its actuators.
Adopted from http://www.rt.el.utwente.nl/agent/
Modeling Adaptive Autonomous Agents, Pattie Maes
ADVIS '04 15
Goal:Goal:An agents goal can take many An agents goal can take many different forms:different forms: End Goals, particular states the End Goals, particular states the agent tries to achieveagent tries to achieve Selective reinforcement or reward that the Selective reinforcement or reward that the
agent attempts to maximizeagent attempts to maximize Internal needs or motivations that the agent Internal needs or motivations that the agent
has to keep within certain viability zones.has to keep within certain viability zones.
Adopted from http://www.rt.el.utwente.nl/agent/
Modeling Adaptive Autonomous Agents, Pattie Maes
Adaptive Autonomous Agents
ADVIS '04 16
Agent
AutonomousAutonomous Capable of effective independent actionCapable of effective independent action
Goal-directedGoal-directed Autonomous actions are directed towards the Autonomous actions are directed towards the
achievement of defined tasksachievement of defined tasks IntelligentIntelligent
Ability to learn and adaptAbility to learn and adapt CooperateCooperate
Cooperate with other agents to perform a taskCooperate with other agents to perform a task
ADVIS '04 17
Agent Types
Cooperate Learn
Autonomous
Collaborative Learning Agents
Smart Agents
Interface agentsCollaborative Agents
ADVIS '04 18
Emergent Phenomena
Emergent phenomena are those in Emergent phenomena are those in which even perfect knowledge and which even perfect knowledge and understanding may give us no understanding may give us no predictive information. In them the predictive information. In them the optimal means of prediction is optimal means of prediction is simulation. simulation. (Vince Darley, 1994)(Vince Darley, 1994)
The whole is greater than the sum of The whole is greater than the sum of the partsthe parts
ADVIS '04 19
Artificial Life Techniques
Agent-based modelingAgent-based modeling Evolutionary programmingEvolutionary programming Genetic algorithmsGenetic algorithms Distributed artificial intelligenceDistributed artificial intelligence Swarm intelligenceSwarm intelligence
ADVIS '04 20
Artificial Problem Solvers:Agent-based Modeling
Computational method where a system is Computational method where a system is modeled as a collection of autonomous modeled as a collection of autonomous decision-making entities that interact in decision-making entities that interact in non-trivial ways.non-trivial ways.
Bottom-up modelingBottom-up modeling Artificial social systems Artificial social systems
ADVIS '04 21
Organizations of agents
Animate agents
Data
Artificial world
Observer
Inanimate agents
If
<cond>
Then
<action1>
Else
<action2>
Courtesy of Lars-Erik Cederman
ADVIS '04 22
Areas of Application
Flow management: evacuation, traffic, Flow management: evacuation, traffic, supermarketsupermarket
Markets: stock market, electronic auctions, Markets: stock market, electronic auctions, ISP marketISP market
Organizations: operational risk, Organizations: operational risk, organizational designorganizational design
Diffusion: diffusion of innovation, adoption Diffusion: diffusion of innovation, adoption dynamics dynamics
ADVIS '04 23
Flow Management
Source: www.helbing.org
ADVIS '04 24
ExposedContracts
DiseaseReports
MoveSpatially
MoveInformation
Agent Location,Demographic
& Social NetworkCharacteristics
Disease Model
Agent Model
DailyCommunity Level
Reports
SharedBSSDatabase
NEDSSCompliant
Geographic Topology
ModelEnvironmental
Lethality
Manifests Symptoms
detectionprivacy
What If ScenarioImpact Analysis
Communication Technology
Model
Courtesy of K. Carley, A. Yahja, B. Kaminsky
Artificial BIOWAR
ADVIS '04 25
Artificial Problem Solvers: Algorithms
Artificial Life tools have led to development Artificial Life tools have led to development of many interesting algorithms that often of many interesting algorithms that often perform better than classical algorithms perform better than classical algorithms within a shorter time. within a shorter time.
These algorithms generally contain explicit These algorithms generally contain explicit or implicit parallelism.or implicit parallelism.
They resort to distributed agents, or to They resort to distributed agents, or to evolutionary algorithms, or often to both.evolutionary algorithms, or often to both.
ADVIS '04 26
Evolving Neural Networks
To develop a hybrid intelligent system – To develop a hybrid intelligent system – Evolving Neural Networks (ENNs) – that Evolving Neural Networks (ENNs) – that can be used in data mining, especially in can be used in data mining, especially in classification problems.classification problems.
ADVIS '04 27
Evolving Neural Networks
Employs computational intelligence Employs computational intelligence methodologiesmethodologies Neural Networks & Genetic AlgorithmsNeural Networks & Genetic Algorithms
Genetic algorithms have been applied to Genetic algorithms have been applied to automatic generation of neural networksautomatic generation of neural networks Feature selectionFeature selection Adaptable topologyAdaptable topology Customized tasksCustomized tasks Ensemble methodEnsemble method
ADVIS '04 28
Optimizing a NN architecture Using GA
Genetic Algorithms
chromosomes
Translation into neuralnetworks
Training neural networks
Evalutation of neuralnetwork performance
f(x)fitness function:
ADVIS '04 29
Feature Selection
Evolving NN 1 Evolving NN 2 Evolving NN n
Features
Final Decision
Combining ModuleGA
GA
Ensemble of ENNs
ADVIS '04 30
Ensemble of ENNs
ENNs meet the major requirements of a ENNs meet the major requirements of a data mining tooldata mining tool Smart architectureSmart architecture
GA GA Self-adaptable structure Self-adaptable structure PerformancePerformance
Ensemble method Ensemble method Accuracy Accuracy Low complexity Low complexity Efficiency Efficiency
User interactionUser interaction Objective function Objective function Customized classification Customized classification
ADVIS '04 31
Artificial Problem Solvers: Reinforcement Learning Methods
Focus on the rational decision-making process Focus on the rational decision-making process under uncertain environmentsunder uncertain environments
Agent can generate a series of actions to Agent can generate a series of actions to influence the evolution of a stochastic dynamic influence the evolution of a stochastic dynamic systemsystem
Underlying control problem is often modeled Underlying control problem is often modeled as a Markov Decision Process (MDP).as a Markov Decision Process (MDP).
ADVIS '04 32
Reinforcement Learning Methods
What to be learned Mapping from situations to actionsMapping from situations to actions Maximizes a scalar reward or reinforcement signalMaximizes a scalar reward or reinforcement signal
Learning Does not need to be told which actions to takeDoes not need to be told which actions to take Must discover which actions yield most reward by Must discover which actions yield most reward by
tryingtrying
ADVIS '04 33
Adaptive Critic Design (ACD)
The neural control design philosophy The neural control design philosophy Algorithms are intermediate between Algorithms are intermediate between
Direct Reinforcement and Value Function Direct Reinforcement and Value Function methods, in that the “critic” learns a value methods, in that the “critic” learns a value function which is then used to update the function which is then used to update the parameters of the “actor”parameters of the “actor”
ADVIS '04 34
Need for Online Hybrid Prediction Model Derived from ACD
Fundamental drawbacks of supervised Fundamental drawbacks of supervised learning-based prediction modellearning-based prediction model
Uncertain volatility in real world call for Uncertain volatility in real world call for adaptive model adaptive model
Reinforcement learning philosophy is Reinforcement learning philosophy is suitable tool especially when the short-suitable tool especially when the short-time performance of forecasting can be time performance of forecasting can be obtainedobtained
ADVIS '04 35
Supervised Learning Assisted Reinforcement Learning Prediction
Architecture for Time-Series
ADVIS '04 38
Artificial Problem Solvers: Robotics
Many robotic systems are currently Many robotic systems are currently being developed in the spirit of artificial being developed in the spirit of artificial life. They are devoted to harvesting, life. They are devoted to harvesting, mining, ecological sampling etc.mining, ecological sampling etc.
ADVIS '04 39
Cooperative Behaviour & path Planning for Autonomous Robots Using Evolutionary
Algorithm & Fuzzy Clustering
ADVIS '04 41
Artificial Problem Solvers: Evolvable Systems
Different categories depending on the Different categories depending on the complexity and purpose:complexity and purpose:
Artificial Life Artificial Life Evolvable Hardware (EHW)Evolvable Hardware (EHW)
analoganalog digital (FPGAs)digital (FPGAs) Hardware design using evolutionHardware design using evolution
Evolutionary Robotics Evolutionary Robotics Evolving controllers for a purpose Evolving controllers for a purpose Co-evolution of robot populationsCo-evolution of robot populations
ADVIS '04 42
Artificial Problem Solvers:Mobile Agents
George Cybenko and Bob Gray Thayer School of Engineering Dartmouth College George Cybenko and Bob Gray Thayer School of Engineering Dartmouth College george.cybenko,robert.gray}@dartmouth.edugeorge.cybenko,robert.gray}@dartmouth.edu
Orders and memos
WirelessNetwork
Technicalspecs
Trooppositions
Wired network
ADVIS '04 43
Static & Mobile Agents Developed for Small Unit Operations
Courtesy of McGrath et alCourtesy of McGrath et al
Objectives:Objectives:• Gather information from sensor reportsGather information from sensor reports• Infer additional information from object ontologyInfer additional information from object ontology• Determine the degree of threat via fuzzy logic inference engineDetermine the degree of threat via fuzzy logic inference engine• Determine recent nearby alerts using clusteringDetermine recent nearby alerts using clustering• Intelligent “push” of relevant threat data via GrapevineIntelligent “push” of relevant threat data via Grapevine
Analysis agentAnalysis agent
Sensor FieldSensor Field
Sensor Report SentSensor Report Sent Threat identified and Alert sentThreat identified and Alert sent
GrapevineGrapevine
ADVIS '04 44
George Cybenko and Bob Gray Thayer School of Engineering Dartmouth College George Cybenko and Bob Gray Thayer School of Engineering Dartmouth College {george.cybenko,robert.gray}@dartmouth.edu{george.cybenko,robert.gray}@dartmouth.edu
Data and simulation cloud
NSF 1998KDIProject
Mobile agentslink weakly
coupleddistributed
components.
Continuous datacollection
Intermittent data
collection
Operational simulation 2
Operational simulation 1
Unexpected (such as emergency relief) uses
Artificial Problem Solvers: Mobile Agents
ADVIS '04 45
Multi Agent Co-operative Area Coverage using GA
Multi Robot SystemMulti Robot System Cover Predetermined Area (Go over every Cover Predetermined Area (Go over every
square inch)square inch) Boundaries MarkedBoundaries Marked Minimize Time and hence Energy EfficientMinimize Time and hence Energy Efficient
ADVIS '04 46
Artificial Problem Solvers: Swarm Intelligence
““Any attempt to design algorithms or distributed Any attempt to design algorithms or distributed problem-solving devices inspired by the collective problem-solving devices inspired by the collective behavior of social insect colonies and other behavior of social insect colonies and other animal societies.“animal societies.“
-[Bonabeau et al., 1999]--[Bonabeau et al., 1999]-
ADVIS '04 47
Swarming Characteristics
Entities share common goal
Local Interaction
s
Self Organizatio
n
Autonomy of units
Stigmergy Simple rules or units
Distributed
Large number or efficient
size
Pulsing of force
Flexible and robust
Swarming
ADVIS '04 49
Ant Colony Optimization
1. Straight Pheromone Trail 2. Obstacle Introduced
3. Two Options are Explored 4. Shortest Path Dominates
ADVIS '04 52
Particle Swarm Optimization
Original intent was to simulate the choreography of a bird flock
Best strategy to find the food is to follow the bird which is nearest to the food
ADVIS '04 54
Particle Swarm Optimization
Global optimum
Courtesy of Maurice Clerk
•The best solution (fitness) particle has achieved so far (pbest)•The best value obtained so far by any particle in the population (gbest)
ADVIS '04 55
Artificial Problem Solvers:Synthetic Ecosystems
The synthetic ecosystems approach The synthetic ecosystems approach applies swarm intelligence to the design of applies swarm intelligence to the design of multi-agent systems.multi-agent systems.
The main concern of research into The main concern of research into synthetic ecosystems is to provide synthetic ecosystems is to provide practical engineering guidelines to design practical engineering guidelines to design systems of industrial strengthsystems of industrial strength
[Parunak, 1997] [Parunak et al., 1998][Parunak, 1997] [Parunak et al., 1998]
ADVIS '04 56
Distributed Architectures for Manufacturing
Holonic SystemsHolonic Systems A whole individual and a part at the same timeA whole individual and a part at the same time ““An autonomous and cooperative building block of a An autonomous and cooperative building block of a
manufacturing system for transforming, transporting, manufacturing system for transforming, transporting, storing and/or validating information and physical storing and/or validating information and physical objects”objects”
[Christensen, 1994][Christensen, 1994] A manufacturing holon comprises a control part and A manufacturing holon comprises a control part and
an optional physical processing part. Multiple holons an optional physical processing part. Multiple holons may dynamically aggregate into a single (higher-level) may dynamically aggregate into a single (higher-level) holon. holon.
ADVIS '04 57
Distributed Architectures for Manufacturing
The application of the holonic concept to The application of the holonic concept to the manufacturing domain is expected to the manufacturing domain is expected to yield systems of autonomous, cooperating yield systems of autonomous, cooperating entities that self-organize to achieve the entities that self-organize to achieve the current production goals.current production goals.
Such systems meet the requirements of Such systems meet the requirements of tomorrow's manufacturing control tomorrow's manufacturing control systems. systems.
ADVIS '04 58
Concluding Remarks
Artificial Life is impacting engineering Artificial Life is impacting engineering systems through Agent-Based systems through Agent-Based architecturesarchitectures
Current Impact Areas:Current Impact Areas: Enterprise Integration and Supply Chain Enterprise Integration and Supply Chain
ManagementManagement Design and Manufacturability AssessmentsDesign and Manufacturability Assessments Enterprise Planning, Scheduling and ControlEnterprise Planning, Scheduling and Control
ADVIS '04 59
Current Impact Areas:Current Impact Areas: Dynamic System ReconfigurationDynamic System Reconfiguration Factory Control ArchitecturesFactory Control Architectures Holonic Manufacturing SystemsHolonic Manufacturing Systems Distributed Dynamic SchedulingDistributed Dynamic Scheduling Commercial scheduling, routing, and force allocation Commercial scheduling, routing, and force allocation
problems problems Use of swarm networks to control swarm Unmanned Use of swarm networks to control swarm Unmanned
Aerial Vehicles (UAV), or undersea vehicles (UGV)Aerial Vehicles (UAV), or undersea vehicles (UGV)
Concluding Remarks