Introduction to Agent-based Modelling


Published on

Published in: Technology
1 Comment
No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Introduction to Agent-based Modelling

  1. 1. Spatio-Temporal Data Mining:Agent-based ModellingEd ManleyDepartment of Civil, Environmental and GeomaticEngineeringUniversity College London
  2. 2. Today’s Talk• Complex Systems• Agent-based Modelling– In Principle– By Example– My Research– The Possibilities• Development of Agent-basedModels– Design Principles– Software
  3. 3. Complex Systems
  4. 4. Complex SystemsSystem Dynamics+ EmergenceAutonomous, decision-making entities“Whole is greater than the sum of it‟s parts”
  5. 5. Complex SystemsNon-Linear BehaviourSystem behaviour is characterised by non-linear actions andinteractionsResponses to actions may bedisproportionate, not easily predicted through examination of macroscopic dynamics only“…increase in price speculationand deliberate hoarding”
  6. 6. Complex SystemsSelf-OrganisationDevelopment of form and pattern through behaviours ofindividuals, with no centralised coordination
  7. 7. Complex SystemsLearning and AdaptationSystems remembers past events and adapts to changeNature of system evolves over time through learning better outcomesArchitectural adaptation in Abyaneh, Iran
  8. 8. Complex SystemsCascading BehaviourCertain behaviours can cascade across networks of individualsBehaviour can transfer quickly between individuals, not allowingthem enough time to adapt to the new conditions
  9. 9. Complex SystemsInter-System ComplexitySystems do not sit in isolation, they interact with others at higher levelsA key challenge for understanding a complex system is identifyingwhere the boundary of influence liesWhat is the cause?Social?Political?Technical?Economic?How far does it influence?City-wide?Nation-wide?World-wide?
  10. 10. Macroscopic phenomena emergethrough Microscopic actions andinteractionsComplex SystemsDriven by Individual BehaviourComplex phenomena are best understood throughconsideration of the behaviour of all interacting parts• How does each individual play a part in the system?• How does individual behaviour change reflect in the system?• How do individuals and systems interact to cause change?• How do interactions vary in respect to other conditions?
  11. 11. Why Study Complex Systems?• Many of today’s most importantglobal challenges are a product ofinteractions between individualsand systems• Social, technological, economic, physical, environmental and politicalsystems all can play a role• Understanding these issuesrequires an insight into thecomplexity of interactions occurringwithin the system
  12. 12. Agent-based ModellingIn Principle
  13. 13. Modelling Complex SystemsTraditional MethodsTraditional approaches – usually incorporating DEs – missout much of the relevant detail, held within the interactionsof individualsHighly non-linear and non-equilibrium processesHigh heterogeneity within the systemIndividuals naturally adapt and change to new conditionsModelling approaches must capture the full extent ofbehaviour, not constrain or smooth them throughmacroscopic equations
  14. 14. Modelling Complex SystemsAgent-based ModellingAgent-based Modelling simulates from individualperspective, capturing emergent phenomena through agentbehaviourIndividual agents are modelled on real-world entities, asautonomous decision-makers that act according to a set ofrulesThe behaviours and interactions of all agents results incollective phenomena reflecting system dynamics
  15. 15. Modelling Complex SystemsWhat is an agent?• A distinct entity, activelyinvolved in dynamics of thesystem• Always autonomous• May behuman, animal, infrastructural, physical, technical, basically anykind of encapsulated individual• Behaviour may be verysimple, or quite complex• May be heterogeneity amongpopulation, but not always
  16. 16. Modelling Complex SystemsAgent-based ModellingAgent-based Modelling enables the exploration of how asystem changes over space and timeIn many cases, spatial proximity is vital, and stronglyinfluences agent behaviour.The behaviour of a system usually changes over time – thisis an explicit component of all agent-based models
  17. 17. Agent-based ModellingIn PrincipleAGENTPropertiesCurrent StatePreferencesGoalsActionsTasksResponsesENVIRONMENTGeographicspace, vectorspace, grid etc.SYSTEMSApply InfluenceEVENT SCHEDULERPrompts all in environment to update statePopulation reflectiveof heterogeneityidentified in realpopulationAgent Behaviour Simulation EnvironmentSYSTEM OUTPUT
  18. 18. Agent-based ModellingModelling Behaviour• Agent behaviour may be simple or sophisticated• However, this does not necessarily reflect in the degreeof complexity viewed at the system level• We’ll now go through some examples of simple andcomplex agent behaviour, demonstrating howaccumulated behaviour reflect in macroscopicphenomena• Models developed in NetLogo software
  19. 19. Agent-based ModellingBy Example
  20. 20. Agent-based ModellingBy ExampleAnt Nest ModelSimple ant foraging behaviourSimple model demonstrating ant communicationbehaviours around food sourcesHeterogeneity across population only lies in agents’actions – yet macroscopic patterns still resolveAnts move randomly in search for food, once foundthey return to the nest, leaving a pheromone trail forother ants to followOther ants, on sensing the pheromone, follow ittowards the food source to carry out the sameprocedureAGENTProperties-ActionsMove RandomlyReturn to NestLeave PheromoneFollow Pheromone
  21. 21. Agent-based ModellingBy ExampleAnt Nest ModelThings to Notice:• Route are established through the initial discovery by one singleant – this is non-linear and self-organisational behaviour• Generally only one route is operating at a given time, with it beingmore difficult to establish other routes while most ants arepreoccupied• Altering the pheromone behaviours causes significant changes torate at which food is found and transported back to nest
  22. 22. Agent-based ModellingBy ExampleSchelling’s Segregation Model (1978)The „first‟ Agent-based ModelModel of evolution of racial segregation in Americancities during 1970sTwo types of simple agents, behaviour leads topatterns developing in urban realmGrid-based spatial representation of environmentSystem outputs showing % unhappy and % similarAGENTPropertiesAgent Type% Similar WantedHappy?Similar Nearby %Other Nearby %ActionsAssess AreaFind Random Spot
  23. 23. Agent-based ModellingBy ExampleSchelling’s Segregation Model (1978)Things to Notice:• Individuals move randomly, assessing each locationindependently, but yet order is established – self-organisation• The random nature of movement means that the arrival of a singlegreen agent may lead to the departure of a number of red agents –non-linear response to a simple action• Space is very important – if space is available agents may clusterinto completely homogenous groups, if not then allowances have tobe made
  24. 24. Agent-based ModellingBy ExampleEl Farol Bar (1994)Exploring Agent Predictive PowerPopular bar in Sante Fe, New Mexico – Thursdaynight is Irish music night! However, the bar often getsuncomfortably crowded putting potential visitors off.Agents use a prediction strategy to decide whetherto go to the bar, prediction is based on a weighting ofprevious weeks’ attendancesAgents utilise their best performing predictionstrategies in deciding whether to attend or not.AGENTPropertiesStrategiesBest StrategyCurrent PredictionAttend?ActionsCreate StrategyMake PredictionGo to Bar
  25. 25. Agent-based ModellingBy ExampleEl Farol Bar (1994)Things to Notice:• Representation of inductive reasoning by agents, bounded in theirknowledge of the problem• Game theoretic representation of behaviour, predicting actions ofothers in order to select own behaviour• Increased number of strategies leads to less variation in patterns• When agents have memory of only previous week’s attendancethen the system descends into a chaotic loop of attendance andnon-attendance
  26. 26. Agent-based ModellingBy ExampleVirus Spreading ModelCascading Virus Spreading through a NetworkAbstract representation of virus spreading through anetwork, demonstrates the SIR(susceptibility, infected, resistance) model ofepidemicsVirus spreads between network nodes, according todefined spread probabilities, with checks carried outevery tickNetwork configuration is vital in determining virusmovement patternsAGENTPropertiesInfected?Resistant?ActionsBecomeSusceptibleBecome InfectedBecome ResistantSpread Virus
  27. 27. Agent-based ModellingBy ExampleVirus Spreading ModelThings to Notice:• Good demonstration of potential for problem to swiftly cascadethrough a network of agents• But, some agents are able to adapt to the virus, causing it toeventually die out• Networked model introduces a new element of systemicbehaviour, demonstrating importance of connectivity• Behaviour of virus – in terms of contagiousness – and agentbehaviour – rate of recovery – are both vital in establishinginfection patterns
  28. 28. Agent-based ModellingBy ExampleEconomic Disparity Land Use ModelSelection of Residence Area by Socio-EconomicStatusMore sophisticated model of land usedevelopment, incorporating four types of agent –representing human behaviour and influence ofinternal systems‘Rich’ and ‘poor’ agents calculate the utility of aresidential location by quality of land measure, cost ofliving and proximity to servicesService location and land squares considered agentstoo – behaving independently, impacting on systemdynamicsAGENTPropertiesTypeActionsCalculate UtilityFind New LocationDieHuman Agents
  29. 29. Agent-based ModellingBy ExampleEconomic Disparity Land Use ModelThings to Notice:• Good demonstration of development of spatio-temporal patternsbased on agent preference and behaviour• Demonstrates mixture of agents and systems and theirinteractions – land value development, location of jobs, selection ofresidence• Behaviour of agents leads to development of quite different urbanstructures – either sprawl or clustering• Development of Schelling’s segregation model, with more complexmechanisms driving dynamics
  30. 30. Agent-based ModellingBy ExampleMy ResearchEvolution of Urban Traffic Systems in response to Unexpected RoadClosuresComplex model of agent behaviour, developed using a moresophisticated modelling platformAgent design incorporates heterogeneity in terms of spatial knowledgeand route choice – based on analysis of movement dataIndividual responses towards road closure and congestion by agentsresults in evolution of traffic patterns over space and time
  31. 31. Agent-based ModellingThe PossibilitiesRecent Conference on ABM included some of the followingthemes:• Pedestrian Simulation, indoor and outdoor• Education Planning• Housing Choice and Housing Policy• Impact of forest composition on spread of disease in Red Colobus populations• Operation of HVAC systems in commercial buildings• Land-use change in the Elbow River watershed in southern Alberta• Agent based model of urban retail system• An Agent-Based Model for Analysing Conflict, Disasters, and Humanitarian Crises in East Africa• Spatio-temporal Dynamics of Slum Formation in Mumbai, India• Climate Change, Land Acquisition, and Household Dynamics in Southern Ethiopia• An Agent-Based Understanding of City Size Distributions• Simulating Residential Dynamics in Londons Housing MarketA world of possibilities!
  32. 32. Development ofAgent-based Models
  33. 33. DevelopmentPrinciples1. Identify the most relevant actors and processes operatingwithin your system of interest– Who is playing an active influence on system dynamics?– What behaviours need to be captured?– Where does the system end? Where are its boundaries?2. Identify whether the relevant behaviours can be modelled– Do methods exist for the modelling of this behaviour?– Can behaviour be measured and quantified? Or reduced into a number ofrules?– Can models reflecting this behaviour be easily validated?
  34. 34. DevelopmentPrinciples3. Over which kind of space and time will your agentsinteract?– Will they move over geographic space, or a vector space?– Will they connect through networks (social, physical) or only byproximity?– Can the space be readily discretised?– How long should the simulation run for? How long a time step?4. Identify the system-level measures that will indicate theeffectiveness of your model5. Select the most appropriate platform on which to buildyour model based on your specification
  35. 35. DevelopmentPrinciples• Keep it Simple!• Greater sophistication does not necessarily lead to abetter model, but will lead to slower processing times• Model only the key agents and key behaviours• Build agent behaviour from data where possible• Validate both agent behaviour and system patterns• Get the design right: Garbage In, Garbage Out!
  36. 36. DevelopmentSoftware• Large number of frameworks, suitability for your workshould depend on the following aspects:– Nature of model (size, complexity, desired output etc.)– Level of programming expertise– Established programming skills– Framework maturity (including access to resources, help)– (Maybe) Operating System• Most available for free (open source), built withinacademia• Many employ design frameworks, scheduling andsimulation engine functionality
  37. 37. DevelopmentSoftwareName Language Maturity NotesNetLogo Proprietary HighEasy-to-use, highly flexible platform;Widely used; Great ‘toy’ modelplatform, but scalability issues.Repast Simphony Java and Groovy HighFlexible, scalable platform; ExcellentGUI with good GIS features; IncludesANN/GA implementations.MASON Java MediumLightweight platform;Less feature-rich than Repast.JADE Java High Standardised development platformSwarm Objective-C MediumSimilar design to Repast but withless functionality
  38. 38. Final Word
  39. 39. Final WordAgent-based Modelling• ABM is an excellent platform for modelling phenomenawhere individual behaviour is central to defining systemcharacteristics• Enables the prediction and examination of expected ANDunexpected behaviours• Applicable to a range of situations, able to incorporatehuman, technical and ecological agents• Careful agent design is vital, and validation should be apriority• The range of platforms available means ABM is open toall
  40. 40. Thank YouQuestions?Ed @EdThink
  41. 41. ReferencesBonabeau, E. 2002. ‘Agent-based modeling: Methods and techniques for simulatinghuman systems’. Proceedings of National Academy of Sciences. 99(3): 7280–7287.Epstein, J. 2009. ‘Modelling to contain pandemics’. Nature 460, 687.The Economist. 2010. ‘Agents on Change’., N. and Terna, P. 2000. ‘How to build and use agent-based models in socialscience’. Mind & Society. 1(1), 57-72.Castle, C.J.E. and Crooks, A.T. 2006. Principles and Concepts of Agent-BasedModelling for Developing Geospatial Simulations., V. and Railsback S. 2012. ‘Agent-Based and Individual-Based Modeling: APractical Introduction’. Princeton University Press.Heppenstall, A,. Crooks, A., See, L. and Batty, M. (eds), ‘Agent-based Models ofGeographical Systems’. Springer.Gilbert, N. 2010. ‘Agent-based models’. Sage.