A Simulation Framework for the Investigation of Adaptive Behaviours in Largely Populated Building Evacuation Scenarios
1. International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008 1
A Simulation Framework for the
Investigation of Adaptive Behaviours in
Largely Populated Building Evacuation
Scenarios
Daniele Gianni, G. Loukas, E. Gelenbe
Intelligent Systems and Networks
Imperial College London
presented by
Daniele Gianni
gianni@imperial.ac.uk
2. 2
Objective
To provide a simulation framework that eases the
investigation of adaptive behaviours in the
context of building evacuation scenarios
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
3. 3
Presentation Overview
Building Evacuation Scenarios
Motivations
Simulated Model
Simulation Framework – Building Evacuation
Simulator (BES)
Incorporating new and possibly adaptive models
Preliminary Validation
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
4. 4
Building Evacuation Scenario
Scenario characteristics:
• Multi-storey building
• Civilians
• Technical Resources in and
Around the Building
• Emergency Event
• Emergency Personnel and
Resources
Civilians try to evacuate
following the quickest and
safest path to the exit, while
adapting to the events
Emergency personnel enter
the building trying to rescue
civilians and extinguish fire
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
5. 5
Motivations
In such scenarios, decentralised optimisation
techniques are needed to support actors to
adapt to unforeseen and changing scenarios
We want to carry out systematic investigations
of such techniques in largely populated
scenarios
Moreover, we want to do it in a cost-effective
manner
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
6. 6
Motivations
Technically speaking, we need a SW framework
that allows:
1. Reproducibility of experiments (systematic
investigations)
2. Extendibility to diverse scenarios (cost-effective)
3. Distributed operation (largely populated scenarios)
i.e. a distributed simulation framework for building
evacuation scenarios that raises the users from
the underlying technical details
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
7. 7
Motivations
The development of such simulators, however, requires
skills and efforts that might not be available or suitable
in particular situations
Skills and knowledge required
• Modelling, parallel programming, theory of simulation, and
simulation infrastructure
Efforts:
• Local / distributed synchronization
• Distributed synchronization and communication
• Implementation of model logic
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
8. 8
Simulated Model
The model includes:
• A world Model, which represents the physical space
inside the building and its status
• One or more hazard agents, which affect the status of
the world
• A population of human agents, which compete and
cooperate for the use of the physical space according
to personal characteristics
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
9. 9
Simulation Framework
The framework is a distributed discrete event simulator
and is based on a layered architecture named SimArch [1]
In this context, SimArch:
• Allows the transparent interchange of the layers’
implementation to fit specific requirements
• Reduces the effort in the development of
distributed simulator by the introduction of
several abstraction layers
Available SimArch implementation on top of IEEE
1516 High Level Architecture standard
[1] D. Gianni, A. D’Ambrogio, and G. Iazeolla, “A Layered Architecture for the Model-driven Development of Distributed Simulators”, The
first International Conference on Simulation Tools and Techniques for Communications, Networks and Systems (SIMUTools08), 3rd
– 7th March, Marseille, France, ACM, 2008.
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
10. 10
SimArch
Distributed Discrete
Event Simulation Layer
Discrete Event Simulation
Service Layer
Simulation Components
Layer
Simulation Model Layer
Layer 0
Layer 1
Layer 2
Layer 3
Layer 4
Distributed Computing
Infrastructure
HLA
DIS
CORBA
WSGrid
General Purpose
Simulation Oriented
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
11. 11
Layer 2: SimJADE
SimJADE is:
• A Java framework for Agent-based M&S
• JADE-based, thus FIPA compliant
It offers a formulation of discrete event
simulation systems in terms of MAS through its
components
It also provides a uniform interface for MAS and
Agent-based M&S, easing therefore the
development of such simulators
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
12. 12
Layer 3: Agents
The agents are defined by independently
specifying:
• A behavioural logic that specifies the interactions with
the external world
• A set of parameters that do not affect the logic
As the principle of “Separation of Concerns”
suggests
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
13. 13
Types of Agents
Three type of agents
• Resource Manager, which manages the
access to the world nodes
• Human agents, which compete for the use of
physical resources
• Hazard agents, which affect the status of the
world
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
14. 14
Human Agents (HAs)
HAs are provided with:
• Personal view of the world
• Models (decision, motion, and health)
• Goal (simple or composite)
HAs present a standard behavioural logic that
bases on movements because they can only
interact with the surrounding world
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
15. 15
Going Distributed
Why? The amount of computational resources
grows at least as polynomial function of the
number of simulated agents
Two major modelling issues to face:
• Model partitioning
• Model adaptation to the distributed environment
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
16. 16
Incorporating New Models
The incorporation of new and possibly
adaptative models is easy and quick as
defining them
These can be decision models (goals and
how to achieve them), motion model and
health model
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
17. 17
Incorporating New Models
For example, we can effortlessly set a graph-
based decision model that computes the
shortest path from the current position to the
exit and which weights are adapted according to
the following definition:
)0.(
)0.(.
)(
fireedgeif
fireedgeifngthphysicalLeedge
edgeweight
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
18. 18
Incorporating New Models
Or, for instance, we can define the motion
model on the nodes as:
Where in the second branch the agent motion is
influenced by the number of queued agents
)0.(.
)0.(
)(
euelengthOfQunodeifnodeTimeeuelengthOfQunodeonstcollisionC
euelengthOfQunodeifnodeTime
nodemotionTime
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
19. 19
Current developments
Distributed Discrete
Event Simulation Layer
Discrete Event Simulation
Service Layer
Simulation Components
Layer
Simulation Model Layer
Layer 0
Layer 1
Layer 2
Layer 3
Layer 4
Distributed Computing
Infrastructure
General Purpose
(CORBA, WS, Globus,
etc.)
Simulation
oriented (DIS,
HLA, ALSP)
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
20. 20
Preliminary Validation
“Absolute” validation is not possible with direct comparison
of data from real world
However, it is still possible in a verifiable scenario. We
considered the following scenario:
• Four floors, three stairwell
• Simple building plan
• 80 employee distributed over the floors (20 on each)
• Since they are employees, it is reasonable to assume that they
know well the floor plan and that are trained to evacuate
Shortest path to evacuate
Orderly and regular motion (FIFO queuing and speed along edges U[1.2;1.5]
m/s) International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
21. 21
Preliminary Validation
40 Experiments using different (independent) sequence
of numbers with good statistical properties and several
civilian distributions over each floor
Average building evacuation time ca 87 s
Simple and optimistic analytical models from Fire
Engineering domain give a total evacuation time in the
same scenario of 76 s
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
22. 22
Preliminary Validation
Behavioural validation
speed on the floor = U[1.6.; 1.8] m/s, speed along the stairs as in the previous experiments
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
24. 24
Conclusions
The use of decentralised optimisation techniques is important to
quickly adapt civilians and emergency personnel behaviour in
building evacuation scenarios
Such techniques, however, require a systematic investigation before
being deployed in real scenarios
Cost and time effective investigations require a software framework
that combines: (1) experiment reproducibility with (2) high level of
extendibility, and (3) distributed operation
We presented the Building Evacuation Simulator, a simulation
framework that meets such requirements, and some basic examples
of use
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
26. 26
Emergency Management
The organization and management of resources and
responsibilities for dealing with all aspects of
emergencies, preparation and on the field (i.e. emergency
scenario)
Two major key aspects in emergency scenarios:
• Actors (i.e. civilians and emergency personnel) have to
continuously and quickly adapt to unpredictable developments
• The behaviour of the individuals affect the final outcome of the
rescue operation
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
27. 27
World Model
A graph models the physical space:
• Nodes: PoI and “collision” points
• Edges: path segments in the simulated world
Graph elements are:
• grouped in regions, which demark the perimeter of perceivable areas
• enriched with a set of attributes that represent the status of the world
on it
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
28. 28
Human Agents (HAs)
HAs are characterised by:
A personal view of the world
One or more goals (including the decision
models on how to achieve them)
Motion model
Health model
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
29. 29
SimJADE components
It is defined through:
A simulation ontology, which defines simulation time and
services and which can encapsulate custom ontologies
A simulation agent society
• Simulation engine (local/distributed), which orchestrates the
simulation
• Simulation entities, which incorporate the logic
An interaction protocol between the agents, implemented
by a set of behaviours and simulation event handlers
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
30. 30
SimJADE Interface
SimJADE maintains a uniform interface with
JADE:
• Hold for a simulation time
• Send a message to another agent at a simulation
time
• Wait for a message
• Conditional wait for a message before a simulation
time
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
31. 31
Resource Manager (RM)
RM is in charge of:
• Coordinating the access to the nodes
• Providing world updates for each agent
RM is defined by a simple wait
event/process event logic that proceeds
until the simulation ends
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
32. 32
Parameters of Human Agents
update()
DecisionModelUpdater
::WorldModel
1*
«interface»
Observable
getMotionTimeOnNode() : Time
getMotionTimeOnEdge() : Time
MotionModel
getLifeTimeOnNode(in n : Node) : Time
getLifeTimeOnEdge(in e : Edge) : Time
wouldDieIn() : Time
willBeAliveAfter(in t : Time)
HealthModel
HumanAgent1*
«interface»
Observer
workTowards() : Node
Goal
1 1
1
*
1
*
getNextNode() : Node
DecisionModel
Node
SimpleGoal CompositeGoal
passToNextGoal()
SequenceOfGoals ConcurrentGoals
workTowards() : Node
recompute()
isAchieved() : Boolean
Goal
1
1
1
*
orderGoals()
GoalSelector
1
1
11
update()
DecisionModelUpdater
1
1
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
33. 33
Logic of Human Agents
Reaching the Node
Waiting for Movement AuthorisationMovement on the Node
Standing on Initial Node
Other Custom DynamicsHA dies
Node
Reached
Ultimate
Goal
Achieved
HA dies
Start Custom
Dynamics
Intermediate Goal
Achieved
Custom
Dynamics
Accomplished
HA dies
Movement
Completed
Movement
Authorised
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
34. 34
Model Partitioning
We follow three guidelines:
• Exploiting the intrinsic parallelism of independent
physical subsystem
• Meeting local memory constraints
• Minimising the network workload
Thus, the simulated world is allocated on
independent single area simulator (floor or
stairs) running on a separate host
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008
35. 35
Model Adaptation
The performance of the simulator are affected
by the amount of data exchanged
Reduce such data by:
• Locally store “constant” data
• Move only individual knowledge
Agents also interact with local world only
Locally, we adopt a condensed representation of
the remote world (GPoI)
International Workshop on Organized and Adaptive Multi-Agent Systems at AAMAS
Estoril, Portugal, May, 2008