Tải bản đầy đủ - 0trang
7 Victory and Defeat Conditions, Scoring and Feedback
allows the user to play the role of the mayor and manage the island for a number
of years, with the mission of ﬁnding an appropriate balance between popularity, economy, attractivity, safety and ecology. SPRITE has a double pedagogical mission: informing the player about a major but often under-estimated risk
(coastal ﬂood); and forcing him to reﬂect on policies for managing this risk. It
provides elected representatives with elements of reﬂection to design an eﬃcient
and balanced management strategy, and allows residents to better understand
(and therefore accept) the policy carried out by their representatives.
The model is fully implemented in GAMA and the game is already playable.
The evaluation of engagement and learning is still preliminary but encouraging.
Short term future work will mainly be dedicated to this evaluation and subsequently improving the model, while longer term prospects include the reﬁnement
of the mechanisms involved in the residents’ decision-making.
Acknowledgements. SPRITE was initially developed by a multidisciplinary
team at the CNRS MAPS 8 thematic school (https://maps.hypotheses.org/
evenements-maps-passes/maps-8) organised by the MAPS network dedicated to multiagent modelling applied to spatial phenomena. SPRITE then served as a basis for the
LittoSim project (directed by Nicolas B´ecu and Marion Amalric) funded by CNRS
D´eﬁ Littoral call 2015, that developed a new, reﬁned, and multiplayer model. Further
reﬁnement of SPRITE (in particular with hydrological and political factors) is ongoing
in the MAGIL project (directed by Eric Barthelemy) funded by CNRS D´eﬁ Littoral
1. Bersani, C., Coll.: Tempˆete Xynthia: retour d’exp´erience, ´evaluation et propositions
d’action. Technical report, Minist´ere ´ecologie, d´ev. durable et mer (2010)
2. Vinet, F., et al.: Le processus de production du risque “submersion marine” en
zone littorale: l’exemple des territoires Xynthia. NOROIS 222, 11–26 (2012)
3. Duvat-Magnan, V.: Les impacts de la tempˆete Xynthia sur les plages de l’ˆıle
d’ol´eron: les r´ealit´es du terrain. Institut du Littoral et de l’Environnement, Technical report (2010)
4. Le Cozannet, G., et al.: Evaluating uncertainties of future marine ﬂooding occurrence as sea-level rises. Environ. Model. Softw. 73, 44–56 (2015)
5. Michael, D., Chen, S.: Serious Games: Games that Educate, Train, and Inform.
Thomson Course Technology, Boston (2006)
6. Sauve, L., Renaud, L., Gauvin, M.: Une analyse des ´ecrits sur les impacts du jeu
sur l’apprentissage. Revue des Sciences de l’Education
33(1), 89–107 (2007)
7. Crovato, S., Pinto, A., Giardullo, P., Mascarello, G., Neresini, F., Ravarotto, L.:
Food safety and young consumers: testing a serious game as a risk communication
tool. Food Control 62, 134–141 (2016)
8. Chadli, A.: Micro jeux et simulation multi-agents participative : apprentissage
des proc´edures de lutte contre les rongeurs arvicoles. Ph.D. thesis, Universit´e des
Sciences et de la Technologie d’Oran (2015)
9. Wouters, P., Spek, E., Oostendorp, H.: Current practices in serious game research:
a review from a learning outcomes perspective. In: Games-based learning advancements for multi-sensory human computer interfaces: techniques and eﬀective practices, pp. 232–250. IGI Global (2009)
C. Adam et al.
10. Grignard, A., Taillandier, P., Gaudou, B., Vo, D.A., Huynh, N.Q., Drogoul, A.:
GAMA 1.6: advancing the art of complex agent-based modeling and simulation.
In: Boella, G., Elkind, E., Savarimuthu, B.T.R., Dignum, F., Purvis, M.K. (eds.)
PRIMA 2013. LNCS (LNAI), vol. 8291, pp. 117–131. Springer, Heidelberg (2013).
11. Brandtzaeg, P.B., Folstad, A., Heim, J.: Enjoyment: lessons from karasek. In:
Blythe, M.A., Overbeeke, K., Monk, A.F., Wright, P.C. (eds.) Funology: From
Usability to Enjoyment. HCI, vol. 3, pp. 55–65. Springer, Rotterdam (2006)
12. Garris, R., Ahlers, R., Driskell, J.E.: Games, motivation, and learning: a research
and practice model. Simul. Gaming 33(4), 441–467 (2002)
13. Koster, R.: A Theory Of Fun For Game Design. Paraglyph Press, Scottsdale (2005)
14. van Ruijven, T.: Serious games as experiments for emergency management
research: a review. In: ISCRAM, May 2011
BDI Modelling and Simulation of Human
Behaviours in Bushfires
Carole Adam1(B) , Geoﬀrey Danet1 , John Thangarajah3 , and Julie Dugdale1,2
Grenoble Alpes University, LIG, 38000 Grenoble, France
University of Agder, Kristiansand, Norway
RMIT University, Melbourne, Australia
Abstract. Each summer in Australia, bushﬁres burn many hectares of
forest, causing deaths, injuries, and destruction of property. Emergency
management strategies rely on expected citizens’ behaviour which diﬀers
from reality. In order to raise their awareness about the real population
behaviour, we want to provide them with a realistic agent-based simulation. The philosophically-grounded BDI architecture provides a very
suitable approach but is little used due to the lack of adapted tools. This
paper uses this case study to illustrate two new tools to ﬁll this gap: the
Tactics Development Framework (TDF) and GAMA BDI architecture.
· BDI architecture · Bushﬁres · Human behaviour
Societies can manage crisis and emergency situations in several ways: adopt
urban and territory planning policies to reduce the risks (e.g. forbid construction
in exposed areas); raise awareness and prepare the population in advance; or
create eﬃcient emergency management policies to deal with crisis when they
happen. Modelling and simulation oﬀer tools to test the eﬀects and complex
interactions of these diﬀerent strategies without waiting for an actual crisis to
happen, without putting human lives at risk, with limited cost, and with a great
degree of control on all conditions and the possibility to reproduce exactly the
same situation as many times as needed.
When modelling human behaviour, mathematical, equation-based models are
too limited ; on the contrary, agent-based models oﬀer many beneﬁts .
They allow to capture emergent phenomena that characterise such complex systems; they provide an intuitive and realistic description of their behaviour; they
are ﬂexible, oﬀering diﬀerent levels of abstraction by varying the complexity of
agents. However, the agents used are often too simplistic, reacting to environmental stimuli without any long-term reasoning. On the contrary, crisis situations
involve complex individual decision making, inﬂuenced by emotions (sometimes
causing irrational actions), and by the social context (eﬀect of group, family).
c Springer International Publishing AG 2016
P. Diaz et al. (Eds.): ISCRAM-med 2016, LNBIP 265, pp. 47–61, 2016.
DOI: 10.1007/978-3-319-47093-1 5
C. Adam et al.
The BDI (belief, desire, intention ) architecture is more sophisticated and
realistic. It describes agent behaviour in terms of mental attitudes, and has also
been used to formalise emotions . BDI provides the perfect level of abstraction to describe human behaviour in terms of folk psychology, which is the preferred level of description for humans . It therefore addresses the problem of
the scarcity of (quantitative) behaviour data by using qualitative data such as
witness statements or expert reports. Despite these advantages making it very
suitable for social simulation, BDI has had limited use in this ﬁeld due to the
lack of adapted tools to harness its complexity . We propose to ﬁll this gap
by introducing to the emergency management community two new tools that
are still under development in the ﬁeld of agent-based modelling and simulation
(ABMS): the TDF methodology that allows the designer to capture informal
descriptions of human behaviour into a conceptual agent-based model; and the
GAMA simulation platform that allows even non-computer scientists to implement a BDI model in an intuitive modelling language and conduct simulations.
We illustrate these tools on a particular case study: modelling the population
behaviour during the bushﬁres that burn every summer across many states in
Australia. Concretely, we focus on the so-called Black Saturday, 7th February
2009, when particularly strong bushﬁres killed 173 people and destroyed hectares
of bush and many properties in the state of Victoria. Reports  showed that
emergency management policies were designed based on an (ideal) expected
behaviour that diﬀered from the residents’ actual behaviour on the day. It is
therefore important to provide deciders with a simulation to raise their awareness
about residents’ decision making, and let them try diﬀerent strategies. For such
a simulation to lay valid results, it is important that the underlying human
behaviour model be as realistic as possible . Currently, the available data
is mostly in the form of witness statements . Given all of this, BDI based
agent-models are an ideal choice.
In previous work , we have shown how the TDF methodology could be
adapted to capture civilians’ behaviour in the ﬁres. We used this methodology to
model the 6 archetypes of behaviours identiﬁed in the population  as 6 possible
roles for the agents of the system, each with their own goals and plans. However,
that report did not provide any statistics about the representation of the proﬁles
in the population, or their possible links with demographics attributes, so it
did not make it possible to initialise a representative simulation with the real
distribution of proﬁles. As a result, in the current paper we adopt a diﬀerent
approach: we develop a single general behaviour for all of the civilians that
captures the diﬀerent ways in which a civilian could behave in the case of a ﬁre.
We then randomly initialise civilian agents with diﬀerent beliefs that will guide
them along diﬀerent paths of the possible behaviours at runtime. We observe and
log the behaviour of the agents, and use these logs to infer the diﬀerent proﬁles
that emerge from them. The proﬁles are therefore not prescribed but observed.
Our hypothesis is that if our general behaviour model of the population is valid,
then we should observe the emergence of the same proﬁles of behaviour as the
ones identiﬁed from the population interviews after the Black Saturday bushﬁres.
BDI Modelling and Simulation of Human Behaviours in Bushﬁres
BDI Modelling Methodology: TDF
In this case study, we aim to model the behaviour of the civilian population in the
Black Saturday bushﬁres, as gathered from witness statements from the aﬀected
population following the bushﬁres . When designing a conceptual model for
computational simulation, UML is the most widely used tool, mostly because
of its generality and ease of use, allowing one to describe entities in terms of
attributes and actions; but it is not well suited for modeling human behaviour
which is what is often required in disaster management and evacuation simulations such as our case study. On the other hand, as mentioned earlier, agentbased software development methodologies that develop systems using mental
attitudes of goals, events, plans, beliefs, capabilities etc. are well suited for these
systems. This is particularly relevant when transcribing behaviours described
by human witnesses as is the case here, since humans naturally tend to explain
their behaviour in terms of mental attitudes. For instance, take this extract: “I
looked out the window and saw some hazy smoke to the north-west. Gary said
that he thought it was just dust but we went outside and straight away we noticed
that we could smell smoke. It was about 12.45pm when we smelt the smoke and
as soon as that happened, Gary agreed to go and get the fire pump”. We can
make the mental attitudes involved more explicit: Gary (wrongly) believed for
a while that the smoke was just dust, but planned to get more information;
after going outside they perceived smoke and realised that it was coming from
a ﬁre (belief update). As a result, he adopted the goal to get ready for the
ﬁre, and started on their plan whose ﬁrst action is to get the ﬁre pump.
Whilst there are several agent-oriented software engineering methodologies
such as Prometheus, Tropos, O-MaSE, GAIA and others  here we introduce a more recent methodology purpose built for eliciting and encoding tactical/strategic behaviour in dynamic domains – TDF (Tactics Development
Framework) [12,13]. TDF is based on the Prometheus methodology , a
mature and popular agent-oriented software engineering methodology. A pilot
study has shown that TDF signiﬁcantly improves comprehension of behavior
models, compared to UML . Although TDF was initially designed to capture
and model military behaviour, we have shown in previous work  that this
framework can be adapted to model civilians’ descriptions of their behaviour in
crisis situations. The TDF methodology proceeds in following 3 phases as relevant to our case study: System specification: Identiﬁcation of system-level artefacts, namely goals, scenarios, percepts, actions, data, actors and roles; Architectural design: Speciﬁcation of the internals of the system, namely the agents that
play the diﬀerent roles, the interactions between the agents (via protocols) if any,
and messages between agents; and Detailed design: Deﬁnition of the internals of
the agents, namely capabilities, plan diagrams and internal messages/sub-goals.
We now illustrate how we modelled our case study in TDF below.
System Specification: Analysis Overview
The purpose of this diagram is to identify the actors (entities external to the
system), the inputs (percepts) to and outputs (actions) from the system, and
C. Adam et al.
identify scenarios (use-cases) that describe possible runs of the system. In our
case, the scenarios are examples of behaviours that we want to observe. A
given scenario will comprise a sequence of steps, where a step could include
goals, actions, percepts and sub-scenarios. Figure 1 illustrates part of the analysis overview diagram related to a scenario where the civilian defends his property
and Fig. 2 illustrates the detailed steps of the “Defend Property” scenario.
System Specification: Goal Overview
In the next step, we develop the goals for the agents (civilians) in the system.
The goal overview diagram illustrates how the high-level goals are decomposed
into more concrete goals. Figure 3 illustrate the goals that civilians adopted,
as extracted from the interviews. The diﬀerent behaviours described in these
interviews result from two high-level goals: Defend Property (stay and defend
it), and Stay Alive (protect life of self and family). The relative priorities of
these goals depend on individual diﬀerences in various factors: awareness of
ﬁre risk, ﬁre training, physical condition, family situation (children to protect),
motivations to defend property (family house, livelihood), etc. These two highlevel goals are then decomposed as shown. The sub-goals of a goal can be either
OR, AND or CON decompositions. OR goals are a disjunction: a civilian can try to
stay alive by either taking cover at home, or by going to a shelter. AND goals are
a (possibly ordered) conjunction: to get to a shelter, one must ﬁrst prepare their
house and themselves, then know a shelter location, then know a safe route to get
there, then ﬁnally follow that route. CON goals are concurrent goals where several
sub-goals must be pursued in parallel: to ﬁght the ﬁre, one must concurrently
monitor their health, the state of their house, and ﬁght the ﬁre (spraying water).
Fig. 1. Partial analysis overview diagram
Fig. 2. Example scenario: civilian defending property (detailed steps)
BDI Modelling and Simulation of Human Behaviours in Bushﬁres
Fig. 3. Goal overview for civilians: decomposition of 2 main high-level goals, “Stay
Alive” and “Defend Property”
System Specification: Role Overview
Having identiﬁed the goals, we next determine the roles to which these goals are
relevant: the 6 proﬁles of behaviour identiﬁed by  from the witness statements.
Usually, these roles will then be assigned to agents, however, in our case we only
model a single agent that can play any of these roles at run-time.
Architectural Design: System Overview
Figure 4 outlines the system overview, which shows the inputs and outputs to the
Civilian agent type. In general, when there are multiple agent types, this diagram
also captures the interactions between the agents via protocols. However, in our
case we have only one (type of) agent.
Detailed Design: Agent Overview
The next step consists in designing the agent overview for each agent type which
details the capabilities and plans of the agents used to achieve their goals. The
plans can be distributed in several capabilities, and each agent can be endowed
with one or more capabilities. Figure 5 shows the diﬀerent capabilities that civilians are endowed with and the inputs and outputs relevant to each capability.
Fig. 4. System overview
C. Adam et al.
Fig. 5. Agent overview, illustrating the capabilities of a Civilian agent
Detailed Design: Capability Overview
Figure 6 details the plans and sub-goals involved in the “defend property” capability. This diagram shows the ﬂow of activities and also speciﬁes the data elements required, for example the “FireData” that is read by the “DefendProperty” and written to by the “ContactFireStation” and “ObserveSurroundings”
plans. Note that this capability contains a sub-capability called “FightFire” that
handles the “ﬁght ﬁre” goal.
Fig. 6. Details of the “DefendProperty” capability
BDI Modelling and Simulation of Human Behaviours in Bushﬁres
Detailed Design: Plan Diagrams
Finally, each plan is detailed in a plan diagram, that is essentially a process
diagram, that can be directly translated into implementation code. A plan can
be triggered by external percepts or by internal goals, and is composed of several
(sequential or concurrent) steps. Steps can be atomic actions (external), activities
(internal processes, e.g. write data), or sub-goals. TDF allows to design one plan
diagram per plan; these diagrams are quite similar to UML activity diagrams.
Figure 7 shows the plan diagram for the “Defend Property” plan. It illustrates
that the plan is triggered by the “FireAlert” percept, and then concurrently
adopts goals to “PrepareProperty” and “Seek Fire Info”. Having adopted the
goals, it monitors the “FireData” information for when the ﬁre is close enough
to adopt the goal to “Fight Fire”.
Figure 8 shows the plan diagram for the “ObserveSurroundings” plan. It illustrates a decision node where the agent considers if it safe outside before choosing
a mode of observation, followed by a merge node where the plan proceeds to the
next step, which is an activity that updates the agent’s ﬁre data.
Fig. 7. DefendPropertyPlan
Fig. 8. ObserveSurroundingsPlan
Simulation Platform: GAMA
GIS and Agent-Based Modelling Architecture (GAMA)
GAMA [15,16] is an open source platform for agent-based modelling and simulation of complex spatialised systems. It provides built-in functions for using Geographical Information Systems (GIS) data, such as OpenStreetMap (OSM) for
fast and precise mapping of the environment. Simulations built with GAMA are
scalable, since the platform can deal with several thousands of agents, depending
C. Adam et al.
on the level of complexity of their architecture. Further, GAMA provides a very
simple and high level programming language called GAML, that allows even
non-programmers to simply build and maintain their own models. As a result,
it is widely used by designers from many diﬀerent ﬁelds. Finally, it is supported
by an active development team that is progressively improving the software.
In particular, GAMA was recently extended with a BDI plugin [17,18] to
allow designers to easily create BDI agent models in the GAML language. They
can specify logical predicates, initialise their agents with beliefs and desires,
describe the eﬀect of new percepts on the agent beliefs, and provide them with
a plan library. The BDI engine then lets the agents perceive their environment,
update their beliefs and desires, select an intention based on relative priorities
of their goals, and choose and execute an adapted plan to reach that goal.
Implementation of the Model
Environment. The BDI implementation for this particular case study is based
on our previous simulation  which used a ﬁnite-state machine (FSM) for
the agent architecture. Only the environment is the same: a square grid of 50
by 50 cells with two safe shelters and a number of houses each inhabited by
one resident, where a ﬁre starts in a few initial cells and then propagates to
neighbouring cells at a speed that can be set as a parameter. The focus of this
paper is not on the realism of the environment or the ﬁre behaviour, but on
illustrating the use of the GAMA BDI plugin to simply design complex human
behaviour models such as that developed using the TDF methodology.
Population. Residents are represented by a GAMA species Civilian. Their
attributes include: a health value which decreases due to injuries caused by the
ﬁre or its radiant heat; a velocity when moving; a random awareness of risks
(perception radius); a random ability to ﬁght ﬁres (defense radius); a level of
determination to protect their property; a random risk aversion; and a reference
to its property. Civilians also maintain their own list of known ﬁres and known
shelters. At the start of the simulation, all agents are unaware of any ﬁres (since
there is none yet) and are waiting at home until they perceive one.
The next paragraphs illustrate how the model designed with TDF precisely
map with concepts provided by the BDI agent architecture of GAMA.
Mapping TDF Design with GAML Code
Predicates. In GAML, the designer ﬁrst needs to describe the diﬀerent logical
predicates that will be manipulated by the agent. This is basically the ontology
of the domain. The code snippet in Fig. 9 illustrates some predicates for our
bushﬁre domain. Predicates can be associated with a priority: here staying alive
has a priority based on the agent’s danger aversion, while protecting the property
has a priority based on the agent’s determination.
BDI Modelling and Simulation of Human Behaviours in Bushﬁres
predicate fire position ← new predicate(’’Know the fire
predicate stay alive ← new predicate(’’Stay alive’’)
with priority rnd(danger aversion);
predicate protect property ← new predicate(’’Protect property’’)
with priority rnd(determination);
Fig. 9. Code snippet: predicates for domain ontology
Percepts. Agents can then be given perceptions, that explain how they interpret
stimuli coming in. These perceptions can add new beliefs or goals. They match
the percepts in TDF that represent the stimuli coming in from the environment
and triggering the agent’s plans. The code snippet in Fig. 10 shows how the
civilian agents perceive new ﬁres (that are not yet in their list of known ﬁres):
they add them to their list, and create a belief that there is a ﬁre, and a desire
to get more information.
perceive target:(list(fire) - known fires) in: perception radius
add self to: myself.known fires;
do add belief(fire position);
do add desire(get information);
Fig. 10. Code snippet: perception of an unknown ﬁre
Actions. Agents are endowed with a number of actions, as speciﬁed in TDF.
There is no concept of capability to group actions together in GAML. The snippet in Fig. 11 illustrates 2 actions of civilians: prepare their building (the eﬀect
value is computed based on some parameters not detailed here, and added to
their building resistance); and prepare themselves (similarly, a value is computed
and added to their total health, to simulate e.g. wearing protective clothes).
Plan Library. Similarly to TDF, actions are combined in plans. The agents are
endowed with a library of plans to achieve their goals. Each GAML plan is
deﬁned with several features: the goal that it achieves (keyword intention); a
context condition (keyword when) that describes when this plan is applicable;
and a success condition (keyword ﬁnished when). The code snippet in Fig. 12