Architecture of a discrete-event and agent-based

36

Int. J. Advanced Intelligence Paradigms, Vol. 4, No. 1, 2012

Architecture of a discrete-event and agent-based crisis response simulation model Rafael A. Gonzalez Systems Engineering Department, Javeriana University, Cra 7 40-62 Bogotá, Colombia E-mail: [email protected] Abstract: This paper presents an architecture of a discrete-event and agent-based crisis response simulation model. In multi-agent systems as a computational organisation, agents are modelled and implemented separately from the environmental model. We follow this perspective and propose an architecture in which the crisis environment is modelled using discrete events and entities, and the crisis response organisation is modelled as a multi-agent system. The combination of both models allows for independent modifications of the response organisation and the scenario, resulting in a testbed that allows experimenting with different coordination mechanisms to respond to the same scenario. In particular, we provide the results of an experimental design where an initial screening shows the impact that different coordination mechanisms have on the overall performance of the response. Keywords: discrete event simulation; agent based simulation; coordination mechanisms; crisis response; design of experiments; DOE; simulation model architecture. Reference to this paper should be made as follows: Gonzalez, R.A. (2012) ‘Architecture of a discrete-event and agent-based crisis response simulation model’, Int. J. Advanced Intelligence Paradigms, Vol. 4, No. 1, pp.36–53. Biographical notes: Rafael A. Gonzalez is a Systems Engineer from Javeriana University (2000) in Bogotá, Colombia. He obtained his MSc in Computer Science from the Delft University of Technology (DUT) in The Netherlands (2003), with a scholarship from the Dutch Government. He obtained his PhD Cum Laude in Systems Engineering from DUT in 2010. He works as Assistant Professor of Javeriana University in the areas of information systems and software engineering, in addition to serving as IT Consultant for projects in the public and private sector.

1

Introduction

Given the difficulties in gathering data, designing controlled experiments and getting access to real emergencies, we often have to rely on recreation of crisis scenarios for training, planning and research in the domain of crisis response. The two main approaches to recreate crisis scenarios are simulations and drills (Massaguer et al., 2006). Agent-based simulation is particularly useful for crisis response, in part because of the ability to quickly refine and extend agent-based models (Robinson and Brown, 2005). Such extensibility allows for addition of new behaviours (Comfort et al., 2004) and new Copyright © 2012 Inderscience Enterprises Ltd.

Architecture of a discrete-event and agent-based crisis response simulation

37

roles which can be integrated to and compared with old ones (Massaguer et al., 2006). More importantly, agent-based simulations, given their capacity to produce aggregate behaviour, can be used to produce complex patterns of interactions which emerge for the analyst or researcher to observe (Comfort et al., 2004). However, agent-based models typically focus on response agents or organisations, e.g., firemen, but are not necessarily appropriate for modelling the crisis situation itself. Other representations may be required for the objects, events and dynamics of the environment in which the agents interact. Discrete-event simulation is process-oriented and events are arranged discretely, which is in line with how crisis simulations (computer-based or not) are typically conceived. For example, a description of a crisis response exercise starts with an incident and further events are added in an historical fashion, as the situation is set to evolve or escalate. Discrete-event modelling is well-suited for the entities and relationships that are affected by these events or which in turn generate new events. For example, environmental entities such as vehicles or infrastructure can affect the dynamics of an evolving fire incident. These environmental entities differ from the response agents in that they do not necessarily require a high degree of autonomy or interaction, at least from the point of view of the response – for instance, vehicles here refers to ‘civilian’ vehicles involved in an accident and not to professional response vehicles, such as ambulances or fire trucks. Accordingly, we suggest combining the benefits of both approaches in an architecture that uses agent-based and discrete-event modelling, integrating them but maintaining enough separation so as to preserve the qualities of each approach and enabling independent changes in both the crisis situation model (discrete-event based) and the response organisation (agent-based). Previous research has already tested the plausibility of combining discrete-event and agent-based simulation. It has been shown that classic problems of discrete-event simulation (DES from now on), e.g., the generic job shop, can be simulated as a multi-agent system (MAS from now on) (Zhou et al., 2008). Other approaches aim at combining the process-oriented approach of DES and the autonomous characteristics of MAS by adding a simulator with artificial time mechanisms to study different design choices enacted by the agents (Janssen and Verbraeck, 2005). On the one hand, it is possible to bring DES concepts into MAS (Gianni, 2008). This option extends agent behaviours to support DES behaviours, such as event handling and notification. On the other hand, agent-based systems can also be built inside a DES environment (Kádár et al., 2005). In this case, communication between the agents is added as an extension to the DES model, using specifically developed agent interaction protocols. The architecture presented in this paper uses the process-oriented, discrete-event characteristics of DES in order to model the crisis environment, that is, the entities and events related to the emergency. The underlying DES simulator handles the events and the random number generation. In addition, we use MAS as an adequate representation of crisis response agents (e.g., firemen) because we want to support their autonomy and be able to model different coordination mechanisms between them. However, we also allow for the agents themselves to be modelled inside the DES environment as entities: animated proxies which handle the physical interaction with the crisis objects (after being prompted to do so by the corresponding agent). This separation between the autonomous agent and its dependent physical proxy, is not strictly necessary, but allows the MAS and DES components of the architecture to evolve separately. Thus, it is possible to produce new crisis scenarios (dynamic descriptions of an incident) relying on the DES

38

R.A. Gonzalez

environment, while at the same time making it possible to add new agents or new agent behaviours to the MAS organisation to alter the way in which the response (and specifically its coordination) is carried out. The DES environment in our simulation model is based on a fictitious emergency scenario designed by the Dutch Ministry of the Interior for training purposes. By ‘scenario’ we understand the description of a particular crisis incident and how it is expected to evolve in time in order to be the basis for experiments. The environment of the scenario contains an excavator doing work on a road in the jurisdiction of a given municipality. The incident starts when a truck, carrying flammable liquid, crashes onto the crane. This prompts the response of fire, police and ambulance services in what is initially a routine situation. Escalation of the incident occurs when the truck catches fire. The incident becomes larger than originally assessed, more response units are needed and a coordinated response is required from multiple disciplines which will setup a commando place incident (CoPI) operational team. Setting up the CoPI team signifies that the response must now be coordinated multi-disciplinarily, which is the focus of our experiments. The incident grows again when the truck explodes, affecting nearby population and creating two main objectives: putting out the fire and rescuing the victims. The experiments use these objectives to measure performance, as well as the coordination cost incurred in achieving them. The rest of this paper is structured as follows. In the next section we present the high-level architecture of the simulation model. Then we zoom into the packages that contain the discrete-event and agent-based components of the architecture. For screening purposes and to show that the model generates structured data for comparing coordination mechanisms, we present a design of experiments (DoE) in Section 4 showing the impact that different coordination configurations have on the overall crisis response performance. In our last section we provide some discussion and conclusions.

2

High-level architecture

The architecture of the simulation model follows the view that a MAS is a computational organisation that interacts with an environment. The visual representation of this initial separation between the MAS organisation and the underlying environment can be seen in Figure 1. In our case, the environment is defined and implemented separately from the MAS and includes the events and entities related to the crisis. The discrete-event simulation handles the crisis events and generates random numbers for each experimental run (determining for instance the initial position of the civilians) so that the same initial configuration can be used to test different coordination setups. The high-level structural architecture of the simulation model is represented through a set of interconnected unified modelling language (UML) packages (represented as folders), shown in Figure 2. The packages contain the classes which implement crisis entities (environmental model) and response agents (agents) respectively, in addition to a visualisation and an ontology package. The reason for the ontology package is that although the architecture is implementation-independent, it does reflect the philosophy behind the FIPA agent specification (IEEE Foundation for Intelligent Physical Agents, 2009), which requires an ontology for inter-agent communication. The visualisation package is used for providing visualisation functionality to the environmental model.

Architecture of a discrete-event and agent-based crisis response simulation Figure 1

39

MAS computational organisation

Organization Agent Agent

Agent

Agent Inter-agent interaction

Agent Agent

Access to the environment

Environment Source: Zambonelli et al. (2003) Figure 2

High-level architecture of the simulation model

Environmental Model

Agents

DES-based entities related to the crisis scenario

MAS-based organizational structure of the response agents

Visualization

Ontology

DES-based classes defining visualization methods for the entities in the Environmental Model.

Static representation of concepts that agents use for representing knowledge and exchanging messages

The environmental model is the package containing the DES entities for the crisis scenario. In our case, the crisis scenario is the training case presented earlier, but the general idea is that it should be the container for any given crisis scenario that is to be modelled. Essentially, this package is where the discrete-event model resides. All environmental entities are modelled as classes (in the object-oriented sense), which can generate or react to discrete events and which are synchronised by the same discrete-event simulator. The environmental model contains the logic of the environmental entities, but interacts with a visualisation package that contains presentation methods for the simulation. The visualisation package contains the classes that define the visualisation behaviour for the entities in the environmental model. This package is added to enable further flexibility in the architecture. While the environmental model contains the entities related to the crisis scenario (e.g., housing, vehicles, victims), the visualisation package defines

40

R.A. Gonzalez

abstract animated or static objects that define the methods for visualisation and movement when running the simulation. This allows the environmental model to be changed or extended without being attached to particular visualisation design decisions. For example, we have chosen visualisation in 2D, but the same environmental model classes could also extend 3D classes (when provided in the visualisation package) without having to do any major changes to the actual definition of the crisis scenario. An important decision on the architecture must be noted here. The response agents also need to be visualised in the simulation, but there is no direct connection between the agents package and the visualisation package. The reason for this is that we chose to model response agents both as entities inside the environmental model and as agents in the Agents package (in a one-to-one fashion). It follows that when a responder exists as an entity inside the environmental model it extends an animated object from the visualisation package. At the same time, this responder entity will be connected to its corresponding agent inside the agents package, but the latter is independent from the underlying visualisation. The agents package contains the organisational structure of the MAS of response agents. In our case, the agents come from the three main disciplines during a typical emergency, such as the one selected as scenario (fire, medical and police services). Again, the agents are connected to the environmental model by having a ‘proxy’ entity representing the physical responder inside the environment. Although it is equally possible to extend the agents so that they, once instantiated, have an actual presence inside the environmental model, we chose to separate the MAS behavioural descriptions inside the agents package, from the discrete-event responders inside the environmental model. For each agent inside the agents package, there will be one responder inside the environmental model. In a way, this constitutes a separation of body and mind of the responder that enables extensions of the behaviour of the agent regardless of whether it will be deployed in a discrete-event simulation environment or not. Conversely, the responder entity inside the environmental model exists independently of whether it is controlled by an agent or not. In our case, the responder entity only acts when prompted by the corresponding agent in the agents package. The ontology package is provided so that the agents comply with the FIPA specifications, according to which the ontology is the language for exchanging, coding and decoding messages expressed as communicative acts. The ontology can also be seen as a static (and empty) representation of the knowledge of the agents, i.e., their mental model. Hence, the arrow connecting the agents package and the ontology package flows from the agents to the ontology, which remains static. The ontology package contains the ontology objects that should map to the environmental model entities, following an agent’s observation. For example, there will be an ontology object for the fire as well as a fire entity in the environmental model. However, in the environmental model the fire will have a state corresponding to the ‘true’ state of the fire (including for example its size and growth rate), while the fire ontology object, once instantiated inside an agent, will contain information on the fire as observed by the agent. This means that there could be additional information inside the ontologically defined fire, such as a perceived risk value; but there can also be inconsistency between the ontological fire object that the agent knows and the real fire entity that the agent observes (for example, a misperception of size). In any case, the objects inside the ontology package do not change, because they are static descriptions of concepts and not instantiated pieces of knowledge. The ontology package is also useful for coding the communication between the agents, allowing for

Architecture of a discrete-event and agent-based crisis response simulation

41

analysis of coordination, which is the purpose of our simulation. Communication is modelled as interaction patterns between the agents, defining the message templates as composition of ontological objects.

3

Discrete-event and agent-based packages

For the discrete-event simulation we have chosen to use the D-SOL discrete-event JAVA-based simulation suite (Jacobs et al., 2002). For the MAS we have chosen JADE as the JAVA-based agent framework (Bellifemine et al., 2008). The fact that the DES and MAS frameworks are both JAVA-based allows for an almost seamless integration of the two without having to alter or extend the original libraries or indeed for creating additional integration layers. In this section we ‘zoom in’ on the packages presented in the previous section presenting the class diagrams in each package and how they are related to the underlying MAS and DES frameworks.

3.1 Discrete-event based packages As noted before, the environmental model is designed as a DES simulation model, where entities are subject to events and dependent on the same time-advancement mechanism. There are entities related to the emergency situation (including physical representations of the response agents); and there are generic classes defining the behaviour of visualisation aspects. Both are contained in the packages shown in Figure 3. Figure 3

Discrete-event based packages of the architecture Environmental Model Civilian

Responder

Model

Vehicle

House

Fire

Visualization AnimatedObject

StaticObject

The environmental model package contains a model class where all the entities are instantiated. This model class implements a DES model interface as defined by the D-SOL simulation suite. This basically provides the means for the model to be constructed with a reference to a DES simulator that is instantiated at the application

42

R.A. Gonzalez

level – the application itself can be managed through a utility class as in our case (not shown in the architecture) or using the interface provided by D-SOL. Once the model is constructed for each simulation experiment, all environmental entities are instantiated in its context with a reference to the same simulator. In this case, those entities are implemented as the following classes: civilian, responder (physical proxy for the response agent), vehicle, house and fire. Given that we need the response agents from the agents package to communicate with the environmental model, we have the Model class extend a JADE agent. This means that the model class is both a DES model and an agent. However, this has no influence on the DES aspects of the Model itself, since the agent-based behaviours are defined separately and will not be active until an agent container (in this case provided by a JADE plug-in for Eclipse, called EJADE: http://dit.unitn.it/~dnguyen/ejade/) has been deployed. Without an agent container and an agent setup, the model simply behaves as a DES simulation model. Because the model is not really an agent inside the agents package – in the sense that it exhibits no autonomous behaviour – it only implements the standard setup and take down methods of a JADE agent. This allows messages to be exchanged between the model and the agents (for example, the first alarm is broadcast by the model to the agents as a response to the crash). The model then acts as a centralised gateway between the agents and the entities of the crisis environment without having to change either the implementation of the environmental model or the implementation of the agents. The fact that both JADE and DSOL are JAVA-based and that the model class implements a D-SOL model and extends a JADE agent, allows for this simultaneous integration and separation of both components of the architecture.

3.2 Agent-based packages The MAS architecture is built in line with the GAIA methodology (Zambonelli et al., 2003). In this approach to analysis and design of MASs, the architecture is equivalent to the organisational structure of the system, itself a combination of the topology and control regime of the agent organisation. The topology in this case follows the basic organisational response structure as defined by the Coordinated Response Procedure of The Netherlands for the Rotterdam-Rijnmond Region in which we base our studies (Trijselaar, 2006). An initial diagram of the topology including the control regime (peer, dependence and control relationships) is shown in Figure 4. At the bottom level are the first responders of the three basic disciplines: firemen, policemen and medics. The operational leadership of these levels is represented respectively by the OvD (Fire Chief Officer), the OvD-P (Police Chief Officer) and the OvD-G (Medical Chief Officer). When an emergency reaches the point in which multidisciplinary response is required, these disciplines must act together, requiring the setup of a multidisciplinary response team called CoPI Team, which requires a leader: the CL (CoPI Leader). In case the emergency escalates to the regional level, the commanders of each of the services need to be involved, establishing a level of control and coordination at the regional level, on top of the already existing CoPI. The commanders from the different disciplines are shown in the figure as CvD (Fire Regional Commander), CvD-P (Police Regional Commander) and CvD-G (Medical Regional Commander). A regional multidisciplinary response team for coordination needs to be setup in an equivalent but higher level than the CoPI: the RegOT (or Regional

Architecture of a discrete-event and agent-based crisis response simulation

43

Operational Team), which is led by the OL (Operational Leader) shown as the uppermost agent in the organisational structure. Figure 4

GAIA-based organisational structure of the MAS OL Depends

Control

Control Peer Peer

CvD

CvD-G

Control

CL

Control

Peer

CvD-P

Control

Depends

Control

Control

Peer Peer

OvD Control

Fireman [1]

Control

Control Peer

Fireman [n]

OvD-P

Peer

Policeman [1]

Peer

Control Peer

Policeman [m]

OvD-G Control

Peer

Medic [1]

Control Peer

Medic [o]

Peer

The organisational structure of the MAS can be simplified by defining more than one role for certain agents. In this case, the CL will by default be the Fire Chief and the OL will be the Fire Regional Commander. In addition, because our study focuses on an emergency that does not scale up to the regional level, we need only consider the CoPI level. Accordingly, the agents package shows the agent classes limited according to the above design decisions, as seen in Figure 5. Having defined the MAS organisational structure, we follow the design by defining a FIPA compliant ontology for the agents to exchange messages (also on Figure 5). The domain ontology in JADE describes the elements that agents use to create the content of messages, specifically concepts, predicates and actions (Bellifemine et al., 2008). Each of these three elements must be represented through a schema and a JADE class implemented for each schema. The empty description of the ontology is defined in the ontology package. The concepts inside the ontology package are the semantic elements of the vocabulary that the agents use, and for this scenario have been defined as: estimated population, fire, infrastructure element (such as a house), location (generic concept used for describing the location of entities in the scenario), material response resource (such as fire truck or traffic control elements), responder (the physical representation of the response agents present in the environment), and traffic (indicating the heaviness of traffic from a specific point of view). Predicates are the structural elements that link concepts together. In our case, we use observation as the predicate that asserts a given observation about the environment by a given agent at a given time. This observation is used as the main container for the mental model of each agent. Actions are special concepts that denote agent actions. In our case alarm is used to request an action from the agents (getting to the incident location).

44

R.A. Gonzalez

Figure 5

Agent-based packages of the architecture Agents ResponseAgent

Fireman

Policeman

Medic

OfficerAgent

FireOfficer

PoliceOfficer

MedicalOfficer

Ontology Alarm

Time

Fire

InfrastructureElement

Traffic Observation

EstimatedPopulation RespondResource

4

Responder Location

Screening through DOE

The experimental design that follows is a systematic way of performing initial experiments with the model and obtaining data on sensitivity of the factors as well as on the impact that different coordination configurations have on system performance. We consider the goals of a simulation experiment to be: 1

understanding a system

2

finding robust solutions

3

comparing two or more systems (Kleijnen et al., 2005).

We will refer to this as the understanding/robustness/comparison view of simulation experiments. This is different from the goals of a traditional experiment, which are: testing hypothesis about factor effects, seeking an optimal policy, or making predictions about performance (Kleijnen et al., 2005). The former goals are more in line with the use of modern simulation models and enable choosing adequate experimental designs, rather than fitting traditional experimental design to simulation. In our case, the goals of understanding and evaluation are aimed at developing coordination theory in crisis response, where there is still theory lacking in terms of emergent coordination.

Architecture of a discrete-event and agent-based crisis response simulation

45

Besides the differences between traditional experiments and simulation experiments, there is one additional contextual factor to be considered in our design: our guiding modelling approach for the crisis response organisation was agent-based. The objective in using agent-based simulation is not prediction or optimisation, because usually there is no possibility of collecting sufficient real world data even to calibrate the agent-based simulation, let alone to use it for credible prediction or optimisation (Sanchez and Lucas, 2002). Agent-based simulations are instead helpful for understanding/robustness/ comparison, especially when there is an interesting tension between the micro (local agent) and macro (global system) levels, as is the case when comparing bottom-up with top-down coordination mechanisms. Statistical DOE can be defined as selecting the combinations of factor levels that will actually be simulated when experimenting with the simulation model (Kleijnen, 1999a, 1999b). Because DOE has been beneficial in real-world settings and in simulation settings that follow the understanding/robustness/comparison view, especially when no real world data is available (Kleijnen, 1999a, 1999b; Kleijnen et al., 2005; Sanchez and Lucas, 2002), this approach will be used for an initial screening of the simulation model that can help in defining more refined experiments to obtain theoretical insight.

4.1 Responses As stated above, agent-based models (and often crisis response simulation models as well) have a large number of factors or variables to be considered. In this section we consider a first screening of those variables in our model, based on previous simulation experiments both in the field of agent-based simulation and crisis response simulation. We begin by listing the dependent variables, or key performance indicators (KPIs), which in DOE are labelled responses: •

Response time: ultimately the efficiency of the response is determined by how long the overall response takes. Indeed, it can be seen as the most critical measurement of performance and its average can be used for performance analysis (Chen and Decker, 2005).

•

Total victims: one of the main goals of crisis response is protecting human life (Suárez et al., 2005). This variable determines in the end how many (or what percentage) of the people involved end up counting as victims. It can be further decomposed according to severity (from minor to fatality).

•

Total coordination cost: the effectiveness of coordination can be measured through coordination cost as a comparative variable. The volume of messages (number of times an agent communicates) can be used to measure coordination cost (Xu et al., 2006b). The assumption is that in MASs “Coordination is achieved through communication by message passing” (Chaudhury et al., 1996). The number of messages can be further decomposed into conversations, defined as a grouping of messages under the same interaction protocol.

4.2 Factors We follow the responses with a list of our independent variables (structural or input parameters), which, in keeping with the DOE terminology, we label factors:

46

R.A. Gonzalez

•

Incident: this factor determines the initial conditions of the incident (location, size, and explosion time). Given the crisis scenario, the exact location cannot be determined directly, but the vehicle that crashes can be determined in advance, shifting the initial location according to vehicle placement. This factor is used similarly in Chen and Decker (2005) and Comfort et al. (2004).

•

Civilians: this factor determines the initial configuration of civilians (their number, location and life function). The life function is defined as the time between being affected by the fire and the time of death without assistance. This means that a shorter life function requires a faster rescue and vice versa. This factor is used in Chen and Decker (2005).

•

Responders: this factor determines the initial configuration of responders (their number per discipline, their starting location, and their profile). This factor is used in Delgado et al. (2003), Excelente-Toledo and Jennings (2003) and Massaguer et al. (2006).

•

Infrastructure: this factor determines the initial number and location of infrastructure elements, simplified in terms of houses and vehicles. Similarly used in Massaguer et al. (2006) and Xu et al. (2006a).

Since the main goal of the simulation study is to evaluate different coordination mechanisms, specifically comparing hierarchically mediated coordination to autonomous mutual adjustment, the key factors related to the MAS are attached to coordination inside and between disciplines. •

Coordination of firemen among rescue or containment: this is mono-disciplinary coordination, where a single resource (a fireman), must be shared between two activities (fight fire or rescue). It can be achieved hierarchically when assignment of firemen between rescue and fire fighting is done through mediation of the fire officer. Alternatively, it can be managed without centralisation when assignment of firemen between rescue and fire fighting is done autonomously by each fireman cyclically. After completing one rescue or one fire fighting cycle, the decision is revised, but still autonomous. Given the choice, this factor can be operationalised as a Boolean variable (either mediated or autonomous).

•

Coordination between firemen and medics for victim rescue: this is a multi-disciplinary coordination dependency, where two activities (rescue by medic or rescue by firemen) should produce a single resource (assistance of a single victim). It can be achieved aided with the mediation of a shared data space between the responders, which is equivalent to the practice of (digitally) tagging patients during an emergency. Alternatively, it is achieved through mutual adjustment by observing the result of others’ actions in which case two or more responders may ‘target’ the same victim and then re-distribute themselves if, upon arrival, the victim has already been assisted.

4.3 Design of experiments The first step in DOE is a factorial design with the parameters that will vary in the experiments. The previous section identified four crisis related factors and two coordination related factors. In order to enable comparison based on the same crisis

Architecture of a discrete-event and agent-based crisis response simulation

47

setting, several crises related factors will be either fixed or random. All others will take initially a low and a high value (which for the coordination factors implies a Boolean value depending on the coordination mechanism). An overview of the factors can be seen in Table 1. Table 1 Factor

Factors in the experimental design Description

Factor levels

-

Incident

Fixed

-

Location of incident

Dependent on vehicle location

-

Initial fire size

Fixed

-

Time of explosion

Fixed

-

Civilians

-

C

Starting number of civilians

Low (–)/High (+)

-

Life function

Fixed

-

Starting location of civilian

Randomly determined

-

Responders

-

-

Starting number of responders

-

F

Starting number of firemen

Low (–)/High (+)

M

Starting number of medics

Low (–)/High (+)

-

Starting location of responder

Fixed

-

Starting location of firemen

Fixed

-

Starting location of medics

Fixed

-

Responder profile

Fixed

-

Firemen Speed

Fixed (faster than medics)

-

Medics Speed

Fixed

-

Infrastructure

Fixed

-

Starting number of houses

Fixed

-

Starting number of vehicles

Fixed

-

Starting location of houses

Fixed

-

Starting location of vehicles

Fixed

A

Fireman assignment

Mediated (–)/Autonomous (+)

R

Victim rescue

Mediated (–)/Autonomous (+)

Using a gridded or factorial design, where 2k design is the simplest (each factor taking on two possible values, as in our case) and mk is the general form (Kleijnen et al., 2005), we obtain a full factorial design of 32 scenarios (treatments), seen in Table 2. The 32 treatments are run for 10 replications each, resulting in a total of 320 experimental runs. Each run generates a log file containing the values for each of the factors and responses, which is the basis for the experimental results and analysis.

48

R.A. Gonzalez

Table 2 Scenario

Design matrix for the 2n (n=5) full factorial design C

F

M

R

A

1

–(50)

–(10)

–(5)

+(Autonomous)

+(Autonomous)

2

–(50)

–(10)

–(5)

+(Autonomous)

–(Mediated)

3

–(50)

–(10)

–(5)

–(Mediated)

+(Autonomous)

4

–(50)

–(10)

–(5)

–(Mediated)

–(Mediated)

5

–(50)

–(10)

+(10)

+(Autonomous)

+(Autonomous)

6

–(50)

–(10)

+(10)

+(Autonomous)

–(Mediated)

7

–(50)

–(10)

+(10)

–(Mediated)

+(Autonomous)

8

–(50)

–(10)

+(10)

–(Mediated)

–(Mediated)

9

–(50)

+(20)

–(5)

+(Autonomous)

+(Autonomous)

10

–(50)

+(20)

–(5)

+(Autonomous)

–(Mediated)

11

–(50)

+(20)

–(5)

–(Mediated)

+(Autonomous)

12

–(50)

+(20)

–(5)

–(Mediated)

–(Mediated)

13

–(50)

+(20)

+(10)

+(Autonomous)

+(Autonomous)

14

–(50)

+(20)

+(10)

+(Autonomous)

–(Mediated)

15

–(50)

+(20)

+(10)

–(Mediated)

+(Autonomous)

16

–(50)

+(20)

+(10)

–(Mediated)

–(Mediated)

17

+(100)

–(10)

–(5)

+(Autonomous)

+(Autonomous)

18

+(100)

–(10)

–(5)

+(Autonomous)

–(Mediated)

19

+(100)

–(10)

–(5)

–(Mediated)

+(Autonomous)

20

+(100)

–(10)

–(5)

–(Mediated)

–(Mediated)

21

+(100)

–(10)

+(10)

+(Autonomous)

+(Autonomous)

22

+(100)

–(10)

+(10)

+(Autonomous)

–(Mediated)

23

+(100)

–(10)

+(10)

–(Mediated)

+(Autonomous)

24

+(100)

–(10)

+(10)

–(Mediated)

–(Mediated)

25

+(100)

+(20)

–(5)

+(Autonomous)

+(Autonomous)

26

+(100)

+(20)

–(5)

+(Autonomous)

–(Mediated)

27

+(100)

+(20)

–(5)

–(Mediated)

+(Autonomous)

28

+(100)

+(20)

–(5)

–(Mediated)

–(Mediated)

29

+(100)

+(20)

+(10)

+(Autonomous)

+(Autonomous)

30

+(100)

+(20)

+(10)

+(Autonomous)

–(Mediated)

31

+(100)

+(20)

+(10)

–(Mediated)

+(Autonomous)

32

+(100)

+(20)

+(10)

–(Mediated)

–(Mediated)

4.4 Graphical analysis of experimental results In order to analyse the results of each of the experiments we use graphical analysis tools that can provide evidence of structure, sensitivity and interesting cases for further, more refined, experiments. The first type of plot is a scatter plot of the individual and mean values of each response against all 32 scenarios. The individual values for each scenario

Architecture of a discrete-event and agent-based crisis response simulation

49

correspond to each of the ten replications, and the mean is the average for those ten replications. On the X-axis we find each of the 32 scenarios defined in Table 2. For clarity, the X-axis grid divides the scenarios in groups of four, indicating the four different coordination setups which are the main focus of the simulation. The scatter plots for response time and coordination cost are shown in Figure 6. Figure 6

Scatter plots of simulation responses (a) Scatter plot of individual and mean response time of 10 replications against 32 scenarios Response Time

Mean Response Time

Response Time (minutes)

300 250 200 150 100 50 0

4

8

12

16

20

24

28

32

Scenario number

(b) Scatter plot of individual and mean number of messages of 10 replications against 32 scenarios

Messages

Messages

Mean Messages

2.300 2.100 1.900 1.700 1.500 1.300 1.100 900 700 500 300 100 0

4

8

12

16

20

24

28

32

Scenario number

The scatter plots show some evidence of structure in the data. Both plots show two halves (a right and a left side) that indicate that starting at scenario 17 the number of civilians is duplicated and the response time and coordination costs increase accordingly. In Figure 6a, for each group of four scenarios, the expectation embedded in the matrix design of Table 2 was that each subsequent combination of coordination mechanisms would result in increasingly faster response times, given the intervention of a mediator. However, this only holds for scenarios 1–4, 5–8, 17–20 and 21–24. The rest of the scenarios show that the second coordination combination (autonomous rescue and

50

R.A. Gonzalez

mediated assignment) has the worst response times. Another interesting finding is that the completely mediated coordination does not always result in the fastest response time. As can be seen in the last two sets of four scenarios (i.e., 25–32), the third combination (mediated rescue, autonomous assignment) is fastest on average. Unsurprisingly, there is a connection between the response time and the coordination cost, shown as total number of messages on Figure 6b. In order to get more information about the relative importance of the factors, we continue with our second graphical analysis using mean plots which simplify the aggregate effect of each of the five factors (the key for each factor is defined in Table 1) on the main responses on average. Clockwise (starting on the top-left corner), Figure 7 shows the mean plots of main effects on a

response time

b

coordination cost

c

victims

d

fatalities (a subset of victims), respectively.

In each case, the low and high values of the factors are averaged over their corresponding half of the scenarios (16 scenarios or 160 replications each). Figure 7

Mean plots of main effects (b) Mean plot of m ain effects on coordination cost

(a) Mean plot of m ain effects on response tim e 950

170

C(+)

R(+)

F(-)

150

M(-)

A(+)

140

120

A(-)

M(+)

130

C(-)

F(+)

R(-)

110

800

600

A(+)

M(-) C(-)

R(-)

(d) Mean plot of m ain effects on num ber of fatalities 6

C(+) F(-)

M(-)

F(+) C(-)

R(+) M(+)

R(-)

A(+) A(-)

Number of fatalities

Number of victims

F(+)

700 650

A(-)

F(-)

750

C(+)

8

4

M(+)

550

10

6

C(+)

850

(c) Mean plot of m ain effects on num ber of victim s 12

R(+)

900 Number of messages

Response Time

160

F(-) R(+)

5

M(-)

A(+)

4

M(+)

A(-)

3 2

C(-)

F(+)

R(-)

From Figure 7, it is clear that increasing the initial number of civilians will increase all response factors. Conversely, increasing the number of firemen results in a reduction of the response factors, meaning a more effective and efficient response. With regards to the medics, there is an interesting finding: Figure 7b shows that despite the positive effect that more medics have on the response, they also increase the coordination cost by requiring multi-disciplinary coordination with the firemen. In terms of the two coordination dependencies, it is apparent from the figures that mediated rescue is significantly better than autonomous rescue distribution between medics and firemen. With regards to assigning the firemen between fire fighting and rescue, while using

Architecture of a discrete-event and agent-based crisis response simulation

51

mediation does improve the response, it does so in a less pronounced manner; however, it significantly decreases the average coordination cost. Finally, fatalities (Figure 7d) is placed next to victims (Figure 7c) to show that there is more impact on fatalities than on victims, particularly in terms of the number of firemen (F) and the rescue coordination (R). This shows that the number of victims is closely related to the initial number of civilians and that it is unlikely for any configuration to reduce this number further, given that it will always take time for the rescuers to arrive. On the other hand, preventing fatalities can be subject of drastic improvements not only in terms of number of rescuers, but also in terms of how they are coordinated.

5

Discussion and conclusions

With the proposed architecture, changes in the discrete-event and agent-based models can be made without having to change the other. In experimental terms, this allows changing the coordination mechanisms in the agent-based crisis response organisation and evaluating their performance for a specific but separate configuration of the discrete-event based crisis scenario. This strategy can be used to try out different configurations in the agent organisation, to add new behaviours in the individual agents and to test the same configurations and behaviours in different crisis scenarios. The relationship between the ontology and the discrete-event environmental model can be seen in the architecture and suggests a mapping between the two, although not on a one to one basis necessarily, since agents can reason about the environment using additional concepts. This soft coupling between the environment and the ontology maintains the separation between the agent-based and discrete-event based models but requires a link at the architectural level between the two. In our case, such a link can be used for experimental purposes because it allows us to analyse the inconsistencies between the different mental models of the agents and also between these and the actual state of the environment. These inconsistencies can be used as basis for designing coordination mechanisms (such as the shared space used earlier) that can overcome information quality deficiencies. In addition, using ontology in the FIPA spirit permits modelling agent interactions that follow FIPA interaction protocols allowing for a standard definition of different coordination mechanisms. For example, we use request for when the officers request individual observations from the responders and propose for when the officers negotiate a shared awareness of the crisis situation. New mechanisms can be defined in the same style, resulting in new interaction protocols that not only reflect coordination mechanisms of real crisis organisations but may also be applied in designing MASs for supporting these organisations. We are currently designing and executing additional experiments that draw on the results of the screening shown in this paper. By refining the values for the factors it is possible to determine with precision the specific conditions under which one coordination configuration outperforms another in one case, while being less effective in another. This can produce theoretical insight useful for training, selection, and support of specific coordination mechanisms for different sizes and stages of a crisis. Finally, some extensions are possible under the current architecture. As seen in this paper, the Medical discipline is currently left designed but not implemented, so this is a natural extension for the future. For the other agents it is possible to extend the behaviours, which are currently modelled as finite state machines. New behaviours or

52

R.A. Gonzalez

revisions can be made not only for optimising behaviour under current conditions but also to add intelligent capabilities so that the agents themselves can learn from their performance and adjust their coordination mechanisms while the crisis is ongoing. In addition, while the environment is modelled through discrete-event entities, given the purpose of the simulation study, it is quite possible or desirable that certain entities in the environment also be modelled as agent. For example, civilians and civilian vehicles are modelled as entities to determine their behaviour in advance as part of the environment over which emergency responders act. Nonetheless, during a real emergency, these ‘entities’ are indeed active and autonomous and do affect the response – in fact, sometimes civilians act as first-responders during the initial stage of an emergency. This would add additional complexity to the study of coordination within the professional crisis response organisation (the objective of the present study), but it would also add more realistic interactions between the responders and the environment.

Acknowledgements This paper is a revised and extended version of: Gonzalez, R.A. (2009) ‘Crisis response simulation combining discrete-event and agent-based modeling’, paper presented at the 6th International ISCRAM Conference (ISCRAM2009).

References Bellifemine, F., Caire, G., Trucco, T. and Rimassa, G. (2008) ‘Jade programmer’s guide: JADE 3.6’, 18 August, available at http://jade.tilab.com/doc/programmersguide.pdf (accessed on 5 March 2009). Chaudhury, A., Deng, P-S. and Rathnam, S. (1996) ‘A computational model of coordination’, IEEE Transactions on Systems, Man and Cybernetics, Part A, Vol. 26, No. 1, pp.132–141. Chen, W. and Decker, K.S. (2005) ‘The analysis of coordination in an information system application – emergency medical services’, in Bresciani, P., Giorgini, P., Henderson-Sellers, B., Low, G. and Winikoff, M. (Eds.): Agent-Oriented Information Systems II, pp.36–51, Springer, Berlin. Comfort, L.K., Ko, K. and Zagorecki, A. (2004) ‘Coordination in rapidly evolving disaster response systems: the role of information’, American Behavioral Scientist, Vol. 48, No. 3, pp.295–313. Delgado, J., Pujol, J.M. and Sangüesa, R. (2003) ‘Emergence of coordination in scale-free networks’, Web Intelligence and Agent Systems, Vol. 1, No. 2, pp.131–138. Excelente-Toledo, C.B. and Jennings, N.R. (2003) ‘Learning when and how to coordinate’, Web Intelligence and Agent Systems, Vol. 1, Nos. 3–4, pp.203–218. Gianni, D. (2008) ‘Bringing discrete event simulation concepts into multi-agent systems’, in Al-Dabass, D., Orsoni, A., Brentnall, A., Abraham, A. and Zobel, R. (Eds): Proceedings of the Tenth International Conference on Computer Modeling and Simulation, UKSIM 2008, pp.186–191, Institute of Electrical and Electronic Engineers, Inc., Cambridge, UK. IEEE Foundation for Intelligent Physical Agents (2009) The Foundation for Intelligent Physical Agents, available at http://www.fipa.org/index.html (accessed on 5 March 2009). Jacobs, P.H.M., Lang, N.A. and Verbraeck, A. (2002) ‘D-SOL: a distributed Java based discrete event simulation architecture’, in Yücesan, E., Chen, C.H. Snowdown, J.L. and Charnes, J.M. (Eds.): Proceedings of the 2002 Winter Simulation Conference, pp.793–800, Institute of Electrical and Electronic Engineers, Inc., San Diego, California.

Architecture of a discrete-event and agent-based crisis response simulation

53

Janssen, M. and Verbraeck, A. (2005) ‘An agent-based simulation testbed for evaluating internet-based matching mechanisms’, Simulation Modelling Practice and Theory, Vol. 13, No. 5, pp.371–388. Kádár, B., Pfeiffer, A. and Monostori, L. (2005) ‘Building agent-based systems in a discrete-event simulation environment’, in Pechoucek, M., Petta, P. and Varga, L.Z. (Eds.): Multi-Agent Systems and Applications IV, pp.595–599, Springer, Berlin. Kleijnen, J.P.C. (1999a) ‘Statistical validation of simulation, including case studies’, in van Dijkum, C., DeTombe, D. and van Kuijk, E. (Eds.): Validation of Simulation Models, pp.112–125, SISWO, Amsterdam. Kleijnen, J.P.C. (1999b) ‘Validation of models: statistical techniques and data availability’, in Farrington, P.A., Nembhard, H.B., Sturrock, D.T. and Evans, G.W. (Eds.): Proceedings of the 1999 Winter Simulation Conference, pp.647–654, Institute of Electrical and Electronic Engineers, Inc., Phoenix, Arizona. Kleijnen, J.P.C., Sanchez, S.M., Lucas, T.W. and Cioppa, T.M. (2005) ‘A user’s guide to the brave new world of designing simulation experiments’, INFORMS Journal on Computing, Vol. 17, No. 3, pp.263–289. Massaguer, D., Balasubramanian, V., Mehrotra, S. and Venkatasubramanian, N. (2006) ‘Multi-agent simulation of disaster response’, in Jennings, N.R., Tambe, M., Ishida, T. and Ramchurn, S.D. (Eds.): Proceedings of the First International AAMAS Workshop on Agent Technology for Disaster Management, pp.124–130, Hakodate, Japan. Robinson, C.D. and Brown, D.E. (2005) ‘First responder information flow simulation: a tool for technology assessment’, in Kuhl, M.E., Steiger, N.M., Armstrong, F.B. and Joines, J.A. (Eds.): Proceedings of the 2005 Winter Simulation Conference, pp.919–925, Institute of Electrical and Electronic Engineers, Inc., Orlando, Florida. Sanchez, S.M. and Lucas, T.W. (2002) ‘Exploring the world of agent-based simulations: simple models, complex analysis’, in Yücesan, E., Chen, C.H., Snowdown, J.L. and Charnes, J.M. (Eds.): Proceedings of the 2002 Winter Simulation Conference, pp.116–126, Institute of Electrical and Electronic Engineers, Inc., San Diego, California. Suárez, S., Collins, J. and López, B. (2005) ‘Improving rescue operations in disasters: approaches about task allocation and re-scheduling’, in Tuson, A. (Ed.): Proceedings of the 24th Annual Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG), London, UK. Trijselaar, J. (2006) Gecoördineerde Regionale Incidentenbestrijdingsprocedure Rotterdam-Rijnmond. Versie 4.2, available at http://www.rhrr.nl/eCache/DEF/1/956.pdf (accessed on 8 April 2008). Xu, Y., Liao, E., Scerri, P., Yu, B., Lewis, M. and Sycara, K. (2006a) ‘Towards flexible coordination of large scale multi-agent teams’, in Scerri, P., Vincent, R. and Mailler, R. (Eds.): Coordination of Large-Scale Multiagent Systems, pp.287–309, Springer, New York. Xu, Y., Scerri, P., Sycara, K. and Lewis, M. (2006b) ‘Comparing market and token-based coordination’, in Nakashima, H., Wellman, M.P., Weiss, G. and Stone, P. (Eds.): Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems, pp.1113–1115, Association for Computing Machinery, Hakodate, Japan. Zambonelli, F., Jennings, N.R. and Wooldridge, M. (2003) ‘Developing multiagent systems: the Gaia methodology’, ACM Transactions on Software Engineering and Methodology (TOSEM), Vol. 12, No. 3, pp.317–370. Zhou, R., Lee, H.P. and Nee, A.Y.C. (2008) ‘Simulating the generic job shop as a multi-agent system’, International Journal of Intelligent Systems Technologies and Applications, Vol. 4, Nos. 1–2, pp.5–33.