1. Crowd Modelling with Situated Cellular Agents Giuseppe Vizzari Department of Computer Science, Systems and Communication University of Milan-Bicocca

2. Outline MABS, crowd modelling and simulation SCA model at a glance SCA crowd modelling methodology –Spatial structure –Active Elements of the Environment and Field Types –Mobile Agents Sample simulation and preliminary results Current and future works

3. Multi Agent Based Simulation Simulation represents a way to exploit a computational model to evaluate designs and plans without actually bringing them into existence in the real world Several situations are characterized by the presence of autonomous entities whose action and interaction determines the evolution of the system Multi-Agent models are particularly suited to represent these situations, and to support the design and implementation of simulators

4. Crowd modelling applications (I) Designer’s decision support –Evacuation situations –Positioning of signs –Malls and shopping centres Support the study of pedestrian behaviour –Envisioning of different behavioural models in realistic environments –Possibility to perform ‘in-machina’ experiments  Need of effective ways to analyse overall system behaviour, but also to present it

5. Crowd modelling applications (II) How to model, design and implement this kind of software system? The image of Scala Square appears courtesy of GeoSim Systems

6. Possible approaches Analytical –May handle very large simulation scenarios –Entities as ‘mere’ numbers –No strong notion of space Cellular Automata based –May handle a large number of entities –Explicit representation of the environment –Entities are homogeneous (they are conceived as particular states of cells) –Extensions to the basic model are often required (e.g. action-at-a-distance) –Complex behaviours require a very large cell state and transition rule Multi-Agent Systems (MAS) based –May handle a smaller number of entities –Entities are clearly separated by the environment –Entities may be heterogeneous –Only a few approaches and models provide a representation of the environment

7. Cellular Automata (I) CA: discrete dynamical systems often described as a counterpart to partial differential equations (continuous) Discrete: space, time and properties have only a finite, countable number of states. Behavior a system is obtained by the interaction of cells according to a transition rule Not to describe a complex system with complex equations, but let the complexity emerge from the interaction of simple individuals following simple rules An example of "macroscopic" dynamics resulting from local interaction is "the wave" in a soccer-stadium Each person reacts only on the "state" of his neighbor(s). If they stand up, he will stand up too, and after a short while, he sits down again Local interaction leads to global dynamics

8. Cellular Automata (II) A regular n-dimensional lattice (n is in most cases of one or two dimensions), where each cell of this lattice has a discrete state –Finite set of cells Finite set of possible states for every cell (uniformity) Cell evolution is specified by a single transition rule –the same for every cell –based on cell neighbours (no action at-a-distance) CA cells evolve synchronously

9. Cellular Automata and crowd modelling Environment  bidimensional lattice of cells Pedestrian  specific state of a cell (e.g. occupied, empty) Movement  generated thanks to the transition rule –an occupied cell becomes empty and an adjacent one, which was previously vacant, becomes occupied Choice of destination cell in a transition generally includes information which is not provided by basic CAs –Benefit-Cost/Gradient: predefined information related to cell desirability –Magnetic Force/Social Force: model the effect of presence of other agents in the environment (attraction/repulsion of crowds)

10. From CA to Situated MAS Entities are reified, separated from the environment –Agents, not just cell states They may have different behaviours –Possibility to integrate several different action deliberation models –Possible heterogeneous system Entities interact by means of mechanisms not necessarily related to underlying cell’s adjacency –Action at a distance is allowed

11. Situated MAS action and interaction Agents are situated –they perceive their context and situation –their behaviour is based on their local point of view –their possibility to interact is influenced by the environment Situated Agents Interaction models –Often inspired by biological systems (e.g. pheromones, computational fields) –Generally provide a modification of the environment, which can be perceived by other entities –But may also provide a direct communication (as for CAs interaction among neighbouring cells)

12. Situated Cellular Agents (SCA) Multi Agent model providing Explicit representation of agents’ environment Interaction model strongly related to agents’ positions in the environment –Among adjacent agents (reaction) –Among distant agents, through field emission- diffusion-perception mechanism Possibility to model heterogeneous agents, with different perceptive capabilities and behaviour Compare T (f  c,t) = true emit(f) react(s,a b,s’)react(s,a c,s’)

13. Situated MAS and crowd modelling Pedestrians  agents Environment  graph, as an abstraction of the actual environmental structure Movement  generated thanks to the field diffusion- perception-action mechanism –Sources of signals (fields): objects, gateways, but also agents –Agents are sensitive to these signals and can be attracted/repelled by them –Possible superposition of different such effects (amplification/contrast) transport(p,q)

14. Sample Application: Lecture hall (I) Single exit

15. Sample Application: Lecture hall (II) Two exits

16. Previous experiences and motivations Some relevant experiences –In traditional case studies –In more complex scenarios Need to specify clearly how to analyze, model and implement these simulations

17. SCA Crowd Modelling Methodology Definition of the MMASS spatial structure Definition of active elements of the environment and field types Definition of mobile agents (types, states, perceptive capabilities and behavioural specification) Definition of monitored parameters and specification of monitoring mechanisms Specific simulation configuration (number, type, position and initial state of mobile agents, other parameters) Abstract scenario specification Computational model for the scenario Experiment-specific parameters

18. Spatial structure of the environment Spatial structure  discrete abstraction of simulation environment Scale of discretization and adjacency  depend on specific scenario (e.g. 50cm sided cells, Von Neumann neighbourhood with some exceptions) Activity supported by software

19. Active Elements of the Environment and Field Types Movement generated thanks to field related effects (attraction/repulsion) Active elements of the environment  sources of signals, reference points –objects which constraint movement –objects that transmit conceptual information (e.g. exit signs or indications)

20. Mobile Agents - States Agents behaviours can be very composite –Segment behavioural specification into several states  attitudes towards movement W G P E Waiting: passengers on the platform waiting for a train Get Off: people on the wagon that have to get off the train Passenger: agent on the train that has no immediate urge to get off Exit: passenger that has got down the train and goes away from the station S Seated: agent seated on a seat of the wagon State Transition W G E P S

21. Mobile Agents – Movement Utility StateExitsDoorsSeatsHandlesPresenceExit press. W-Attract (2)--Repel (3)Repel (1) P--Attract (1)Attract (2)Repel (3)Repel (2) G-Attract (1)--Repel (2)- S-Attract (1)---- EAttract (2)---Repel (2)- Movement generated thanks to field related effects (attraction/repulsion) When multiple fields are present in the environment –Agents evaluate the utility of each possible destination site –The single contributions of various fields are combined in the overall site utility, for the current agent state transport(p,q) 

22. Case Study Preliminary results Simulation configuration –6 agents getting off –8 agents getting on “Mixed” results –Agents accomplish their goals –Some ‘erratic’ phenomena Oscillations (“forth and back” movements) Semi static situations (“equivalent” groups facing each other)

23. Further works Tuning of utility function –Already some work done Mechanism for agent facing and penalization of steps back choices Penalization of immobility (unless a goal is reached)... define exceptions to crowd repulsion, to support imitation? Integration with realistic 3D visualization tools –Already some results  3D Studio Max integration –Full 3D engine integration in progess Design and development of further tools supporting the modeller in the implementation of SCA crowd simulations Towards more complex behaviours?

24. To obtain a machine-readable spatial abstraction in a semi-automatic way from existing models of the environment To obtain an effective visualization of simulation dynamics Possibly to exploit the rich information of a 3D model to implement highly realistic perception Different possibilities to develop such a system –Integrate results of bidimensional simulators with existing 3D modelling and rendering tools (offline animation) –Integrate the simulator with a realtime 3D engine (interactive simulation) Why 3D? And how? The image of Scala Square appears courtesy of GeoSim Systems

25. Integrating 2D and 3D Java based bidimensional simulator implementing MAS model elements Exported log of the simulation including –Definition of the spatial structure –System dynamics MaxScript that allows 3D Studio Max to generate an animation representing the simulated scenario Avatar#001#001#001#004#003#000@ Avatar#002#002#001#003#005#000@ Avatar#001#002#010#003#002#000@ Avatar#002#001#001#003#004#000@ Space#001#001#001#001#006#000@ Space#001#002#001#002#005#000@ Space#001#003#001#004#005#000@ Space#001#004#001#004#004#000@ Space#001#005#001#003#004#000@ Space#001#006#001#003#002#000@ Space#001#007#001#004#002#000@ Space#001#008#001#005#002#000@

26. “Full” 3D solutions Enhance a realtime 3D engine with MAS model elements Endow visually rich entities with a flexible, general behavioural model Supports –dynamism in visualization –interactivity of the simulation system –exploitation of 3D model of the environment +

27. More complex behaviours? Using the head-body metaphor for agent-based systems, situated MASs provide a good model for the “body” –notions of locality and perception –situated action and interaction Possibility to enhance this body with some more ‘head’ functionalities –From simple reactive models to more complex agent architectures –Internal abstractions of agents’ environment (e.g. maps) Build more complex, psyco/sociologically founded behavioural models on these basic mechanisms Compare T (f  c,t) = true emit(f)

28. Pervasive Computing and Agents? Human-machine interaction is pervading the environment –ever growing number of diffused computational devices –used in every-day activities –influencing human interaction Technology supporting context awareness –provides services both context and location aware –the range of high level applications is ever extending –the challenging applications in the future will be pervading the environment and its inhabitants Distributed/diffused/pervading technology main requirements: –be able to fuitfully exploit interactions with other components (e.g. information sources) –offering reasonable services autonomously

29. From Simulation to Pervasive Computing From a simulation in which synthetic entities interact in a virtual museum To an augmented reality in which actual artifacts provide information and services to museum visitors Situatedness, locality, situated interactions supporting context- awareness

30. Giuseppe Vizzari Artificial Intelligence Laboratory L.INT.AR Department of Computer Science, Systems and Communication University of Milan-Bicocca giuseppe.vizzari@disco.unimib.it