MAGS, a Multi-Agent Geosimulation Platform and its Application to the Simulation of Crowd Behavior, Fire and Gaz Propagation and Road Traffic

Transcription

1 MAGS, a Multi-Agent Geosimulation Platform and its Application to the Simulation of Crowd Behavior, Fire and Gaz Propagation and Road Traffic Dr. Bernard Moulin Cognitive Informatics Group Computer Science Department and Centre for Research in Geomatics Laval University, Québec, Canada December 2004

2 This project is currently financed by Geoide, the Canadian Network of Centers of Excellence in Geomatics and RDDC Valcartier (Team: Y. Van Chestein, D. Gouin, M. Pigeon, D. Thibault + M. Bélanger) Other partners: Alberta Sustainable Resource and Development, Quebec SOPFEU, Center for Spatial Analysis (Mc Master Univ.), CRAD (Univ. Laval), Ville de Québec (Service de police), Sûreté du Québec, Ministère des transports du Québec 2

3 Overview Project history, objective & requirements ( ) Key concepts and techniques MAGS Main Components Creation of a Virtual World and Outputs Agent s Perception, Memory and Navigation Agent s Needs, Goals and Bahaviors Groups of Agents The smoke simulator The scenario editor The traffic simulator Other Projects related to MAGS The MUSCAMAGS Project ( ) Conclusion 3

4 The MAGS Project Project History, Objective, Requirements and Application Domains 4

5 Project history : development of PADI, a tool for the design of parks and the simulation of their usage (financed by GEOIDE Project DEC 30) 2001: Feasibility study for the extension of PADI to simulate crowd behaviors in urban environments (financed by CRDV) : Development of MAGS as a general geosimulation platform based on agents. Application to the simulation of crowd in urban areas and consumer behaviors in shopping malls (financed by GEOIDE and RDDC Valcartier + other sources) 5

6 The MAGS Team (Laval Univ. Computer Science Department) Bernard Moulin (Project Leader) Walid Ali (PhD Jan 2006) Nabil Sahli (PhD Jan 2006) Walid Chaker (Research Assistant , PhD in Progress) Jimmy Perron (MSc 2003, Research Assistant ) Jimmy Hogan (Research Assistant ) Patrick Pelletier (MSc 2003) Fouad Belafkir (MSc 2004) Azzedine Kabli (MSc 2005) Mondher Bouden (MSc 2004) Frédérick Legault (MSc 2003) 6

7 The MAGS team (July 2003) 7

8 The MAGS Project MAGS : Multi-Agent Geo-Simulation General Objective: To develop a software platform that allows the creation of virtual geographic spaces (in 2D and 3D) in which several thousands agents can navigate for different purposes MAGS is written in C++ and runs on Windows platforms 8

9 Requirements for MAGS (1/2) We need to create a virtual geographic environment (VGE) in 2D (and possibly 3D) from reliable sources (GIS) We need to create agents of various types, each agent instance being individualized An agent must be able to perceive its environment and to communicate with other agents An agent needs a memory capability in order to organize its knowledge about the VGE gained from past experience An agent must be able to navigate autonomously and to react to changes occurring in the VGE The agents characteristics must reflect various possible states (static, dynamic, possession, etc.) 9

10 Requirements for MAGS (2/2) The agents behaviors must offer efficient planning capabilities (reactive planning based on objectives) Agents must be able to display group behaviors Agents must be able to react to smoke and dense gaz Scenarios must be specified in an easy way, including the initialization of agents and objects, the introduction of specific events that influence the simulation Different functions (such as trace functions) must be provided in order to analyze the results of the simulations The system must be optimized and allow simulations involving several thousands agents 10

11 Some Application Domains Simulation of crowd behavior in urban environments Simulation of emergency plans involving a large number of agents in urban areas, buildings, etc. Simulation of customer behaviors in shopping malls and in large stores Simulation of smoke and dense gaz and the reaction of a crowd Simulation of the interactions of different groups of agents (police & demonstrators, pedestrians & cars) Scenario building (course of actions) and simulation based assessment using MAGS Multi-agent coordination and planning in dynamic geographic environments (ex. Fire fighters) 11

12 MAGS can simulate the movements of thousands agents in a 3D virtual space 12

13 Interaction of Pedestrians and Vehicles 13

14 Simulation of customer behavior in shopping malls 14

15 Crowd and Smoke 15

16 Key Concepts and Techniques used in MAGS 16

17 MAGS main components VGE Preparation Suite and VGE files (2D and 3D environments) The 3D Characters Preparation Suite The agent knowledge bases The agent knowledge bases management system The multi-agent simulator The scenario editor The traffic simulator The smoke simulator 17

18 Preparation suite Initialisation files MAGS Designer Agents??? Knowledge Bases management Agent s??? Knowledge Bases Agents??? Behaviour Perception Memory Navigation Scenario editor Scenario files Thread Manager VGE Preparation suite 3D Characters Preparation suite VGE files 3D characters files Environment Graphic Engine Interface User 3D specialist Particle System Network AIMSUN Traffic control 18

19 Creation of a virtual world Data sources Geographic Database (GIS) Use of GeoMedia programming utilities Most of GIS formats are supported Possibility to combine data from different sources GIS Data of Quebec city in GeoMedia (ArcView format) 19

20 Creation of a virtual world: outputs (1/3) The Height Map is required by the navigation and perception modules Lower elevation point Higher elevation point Ariadne Map (Streets and paths are colored in red. They are used to facilitate the agent s navigation) These maps are automatically generated from the GIS data 20

21 Creation of a virtual world: outputs (2/3) Building Map associates points to objects and (combined with the properties file) provides additionnal information about the environment Id : 79 Type : House Properties File Position : (403,203) Height : 8 meters 21

22 Creation of a virtual world: outputs (3/3) Zone Map (Zones defined in GIS & automatically generated) Visualization maps Id : 81 Properties File Type : District Ste-foy Position : (603,302) Properties : 2D Model 3D Model 22

23 Agent s main characteristics States (Static, Dynamic, Communication, Possession) Profile and roles Basic behaviors (navigation in the VGE and perception) Needs and objectives (sub-objectives) Behaviors related to the achievement of objectives (actions with activation rules and completion rules) Management of the objectives priorities and temporal constraints 23

24 Agent s perception 3D perception of the environment Vision field (ex. 90 degrees) Computation of the lines of sight using the Height map h= 10 h= 20 h= Building height A Perception approximation method Height Map 24

25 Modification of the agent s perception cone when crossing a road 25

26 Agent s s Memory Environment Perception Perceptual Memory Rehearsal and Retention Filter Working Memory Recall Update Long Term Memory lost Lost Retention 26

27 Agent navigation (1/2) Basic notions Use of the Height Map to represent the navigational constraints : obstacles, slopes, etc. The Height Map provides difficulty degrees for the agent s movements Use of the Ariadne Map to represent paths and trails in order to optimize the navigation and make it more realistic Use of a Location map to represent the positions of all the fixed and mobile agents To compute its next displacement the agent only uses the information accessible in its visual field 27

28 Agent navigation (2/2) Illustration : ray tracing function Destination 1 Ariane Thread Ok Obstacle A : «following a path» B : from «following a path» to obstacle avoidance Agent s current Position C : «obstacle avoidance» 28

29 Behavior management Needs Elementary actions Objectives Complex actions Priority management 29

30 Needs An agent has a set of needs related to its profile(s) Needs are dynamic variable (value change over time) Each need is associated with a priority which is different from one profile to another An objective is related to a need : this is how we determine the base priority of an objective Elementary actions Elementary actions trigger routines modifying an agent s state or an objective s status For example: increase its speed; set the destination target; decrease the agent s hunger while it eats, etc. 30

31 Objectives Objectives are organized in a tree structure Node : Composite objective Leaf : Elementary objective (contains actions) The tree structure does not constrain the execution order of the objectives A tree defines a plan for an agent Activation rules Completion rules Priority of objectives 31

32 MAGS 2002 version: Example of buskers 32

33 Complete structure of an objective Elementary objective 1.1 : To have found a presentation spot Activation Rule : There is no precondition, the objective will be active when its super-objective becomes active Action : Search a zone for a presentation spot Completion Rule : If a presentation spot is perceived and this presentation spot is free, then make the spot not free, move around the spot and the objective 1.1 is completed with success Composite objective 1.0 : To have performed a street play Activation Rule : If the need for money is greater than a threshold Completion Rule : If the objective 1.4 is completed with success, then the objective 1.0 is completed with success Elementary objective 1.2 : To have attracted a crowd of spectators Activation Rule : If the objective 1.1 is completed with success and the objective 1.3 isn't active, ongoing or completed with success Action : Send message to gather spectators Completion Rule 1 : If there are enough (5) spectators around, then the objective 1.2 is completed with success Completion Rule 2 : If there are not enough spectators around, then the objective 1.2 is completed without success Elementary objective 1.4 : To have collected money Activation Rule : If the objective 1.3 is completed with success Action : Send message to ask spectators to pay for the show Completion Rule : If there is no spectator around, then the need for money is reduced, make the presentation spot free and the objective 1.4 is completed with success Elementary objective 1.3 : To have performed a show Activation Rule : If the objective 1.2 is completed with success Action : Perform a show Completion Rule : If the show was performed for its predefined duration time, then the objective 1.3 is completed with success 33

34 Groups of agents Social group Group of agents sharing one or many common characteristics (i.e. : police officers, students, ) A group is considered as an agent and can have its own behavior (for example : an objective can be used to manage the group members) Cohesive groups (Clusters) Group of agents occupying a restricted closed or open area (flocks, queues, herds, ) Clusters are created dynamically A cluster can have its own behavior (for example : an objective of a queue cluster can manage the movements of the agents in the queue) 34

35 Modeling agents in 3D Use of 3D Studio Max and character libraries Individual agents Different models can be applied on agents Agents are basically animated Cameras We can have a first person camera for each agent in the simulation as well as a third person camera 35

36 Simulation of smoke and gas We simulate smoke (or a dense gas) in a 3D model of space and to integrate these phenomena in MAGS and to make it possible for MAGS agents to react to them We use an approach based on particle systems Wind 36

37 37

38 Scenario editor The scenario editor allows users to assemble scenarios to carry out different simulations, it supports five main stages for scenario specification : - Specification of the environment - Creation of the agents and the objects - Specification of the population of the scenario - Description of the scenario - Creation of agent prototypes The current version of the Scenario Editor implemented in Visual Basic A new version is developed in C++ 38

39 Specification of the environment - The system loads agents included in the environment (buildings, trees, etc. ) which are provided by the GIS with all their properties (position, type, etc.) 39

40 Population Specification - Instantiate agents based on 4 types of agents: Agent, Object, Group, Cohesive Group (cluster) - Assign behaviors to agents using pre-defined prototypes - Initialize agents static and dynamic states 40

41 Scenario Specification - Specify when and where agents will appear - Assign specific goals to agents depending on specific conditions - Specify events that can occur during the simulation under certain conditions 41

42 Virtual Environment Modification 42

43 Modifying the Virtual Enviroment During the simulation it is possible to add, modify and delete any object which is recorded with the.x format (DirectX) Application: the user can modify the environment and observe how agents react to the changes See next slide Example : Trees planted during the simulation 43

44 ADDING A FENCE IN THE ENVIRONMENT A A pedestrian group avoided the fence B A group of pedestrians blocked by the fence E C A group of pedestrians try to go around the fence A B F D D The fence (obstacle) has been added during the simulation. E Point that the crowd attempts to reach C F Group of pedestrians who reached the goal G General direction of movement of the crowd (peace walk setting) G 44

45 Vehicle Traffic Simulation and MAGS 45

46 Simulation of vehicle movements We aim at integrating in the MAGS environment other types of mobile agents (Vehicles, etc.) and to study the spatial interactions between different mobile agents (ex. vehicles and pedestrians) We use a library to simulate vehicles micro displacements: the GETRAM platform 3 main components : TEDI: used to draw roads in a 2D environment AIMSUN: the simulation engine to simulate vehicle displacements GETRAM extension: API used to access the simulation engine 46

47 AIMSUN Simulation based on data created using TEDI Management of displacements for each vehicle: lane change, gap with respect to other vehicles, collision avoidance, management of traffic lights - Real-time simulation (using time steps) - taking into account terrain geometry 47

48 Adaptation of MAGS to use AIMSUN -Automatic generation of road segments using TEDI s data - Addition of vector data (nodes, road segments) - Creation of vehicle agents (with a drive state which is used to update the vehicle position in MAGS) - MAGS and AIMSUN can be on the same or different machines - MAGS agents can generate incidents which are transmitted to AIMSUN and can influence car behaviors (ex. pedestrian crossing a lane) 48

49 AIMSUN MAGS 49

50 Interactions between Cars and Pedestrians 50

51 Other Projects related to MAGS 51

52 Sumo simulation of traffic flow in Quebec city A Network composed of Edges Thousands of vehicles 52

53 Wild Fire Simulation using the Prometheus Library from ASRD (Alberta Sustainable Resource and Development) 53

54 FIRE-MAGS (current implementation) 54

55 Fire Propagation in MAGS 55

56 A New Project: Critiquing System & MAGS Critiquing systems (TIF Project, M. Bélanger) Project Objective: to develop a multi-agent geosimulation system (MAGS) component to support: 1) the course of action (COA) specification / modification, 2) the critiquing system; 3) the exploration and comparison of various scenarios - The MAGS component would be able to get accurate data about terrain, resources, environmental conditions, etc. from GIS and other data bases - Domain specialists (DS) could specify different scenarios (COA) and specify intermediate and final goals (states) - For each scenario, the MAGS component would launch a simulation that the DS would observe - The outcomes of the simulations would be used to assess the strengths and weaknesses of the proposed scenarios 56

57 The Future of MAGS: The MUSCAMAGS Project

58 The MUSCAMAGS Project Multi-Scale MAGS A project submitted to GEOIDE (Accepted in December 2003): duration: April March researchers Laval Univ.: Bédard, Moulin, Thériault, Wilfried Laurier Univ.: Doherty, McMaster Univ.: Scott, Queens Univ: Harrap. Current partners: RDDC Valcartier, Alberta Sustainable Resource and Development, Quebec SOPFEU, Center for Spatial Analysis (Mc Master Univ.), CRAD (Univ. Laval), Ville de Québec (Service de police), Sureté du Québec, Ministère des transports du Québec 58

59 Multi-Actor Dynamic Spatial Situations MADSSs involve a large number of actors of different types (human, animal, etc.) acting in geographic spaces of various extents MADSSs need to be monitored to insure : human security and equipment preservation (flood, earthquake, wildfire, oil slicks), the respect of public order (population evacuation, crowd monitoring and control, peace-keeping, etc.) the adequate use of infrastructures (monitoring of people and households transportation and shopping habits in a urban area to better plan transportation infrastructures, location of services and retailers, etc.) Impact of emergency response plans 59

60 MADSS (continued) Certain MADSSs occur on a regular basis (ex. daily traffic patterns in a urban area) whereas other MADSSs can evolve rapidly as a consequence of the occurrence of particular events and/or changes in individual behaviors (often in crisis situations) Certain MADSSs may occur within the context of another MADSS of larger extent Decision makers need an overall understanding of the situation to monitor its evolution, to develop strategies to adequately intervene, to develop and compare alternative intervention scenarios and to anticipate the consequences of these interventions 60

61 Several issues must be tackled at once : construction of multi-resolution world models with reliable physical terrain models and contextual semantic attribution population data and physical data (ex. fire spreading) monitoring of the situation evolution assessment of different courses of actions anticipation of the reactions of the actors involved in the situation MUSCAMAGS Global Objective To develop a methodology and a software platform to create multi-scale multi-agent geosimulations to support operational decison support systems for MADSSs 61

62 62

63 Application domains RDDC Valcartier and Service de Police de Québec and Sûreté du Québec: Simulation of crowd movements and impact on different activities (ex. impact on emergency evacuation of people) CRAD and associated researchers in Toronto area: Simulation of the travel, shopping and leisure behaviors of people in large urban areas Alberta SRD and SOPFEU: simulation of different strategies when fighting a wild fire (teams of firemen, use of the Prometheus library) Other domains (Shopping malls, environmental health ) 63

64 Questions and Discussion 64