RFID Reader Networks: Channel allocation algorithms, performance evaluation and simulator
A Multi-Agent Simulator for Teaching Police Allocation
-
Upload
villadolores -
Category
Documents
-
view
217 -
download
0
Transcript of A Multi-Agent Simulator for Teaching Police Allocation
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
1/12
I This article describes the ExpertCop tutorial sys-
tem, a simulator of crime in an urban region. In Ex-
pertCop, the students (police officers) configure
and allocate an available police force according to
a selected geographic region and then interact with
the simulation. The student interprets the results
with the help of an intelligent tutor, the pedagog-
ical agent, observing how crime behaves in the
presence of the allocated preventive policing. The
interaction between domain agents representing
social entities as criminals and police teams drives
the simulation. ExpertCop induces students to re-
flect on resource allocation. The pedagogical agent
implements interaction strategies between the stu-
dent and the geosimulator, designed to make sim-
ulated phenomena better understood. In particu-
lar, the agent uses a machine-learning algorithm to
identify patterns in simulation data and to formu-
late questions to the student about these patterns.
Simulation aims to represent one phenom-enon by means of another and it is useful
to measure, demonstrate, test, evaluate,predict, and decrease risks and costs. Practicalapplication can be seen in various areas, suchas in the aeronautical industry, nuclear indus-try, and the military. In educational terms, sim-ulation is important because it allows learningthrough the possibility of doing. Simulationhas been shown to be a good teaching tool, es-pecially for complex situations with high costand risk.
Multiagent systems have been widely adopt-
ed in the development of complex systems.One of the most important reasons to use a
multiagent systems paradigm is for handling
the interaction of different entities or organi-zations with different (possibly conflicting)
goals and proprietary information. A multia-
gent system is also appropriate whenever thereis a need for representing individually each en-
tity of the modeled domain, or if such entities
have an intelligent behavior to be modeled.
Social or urban environments are dynamicand nonlinear and are made up of a great num-
ber of variables, characterizing a complex sys-
tem. The use of multiagent systems to simulatesocial environments has become broadly used
(Khuwaja, Desmarais, and Cheng 1996). Aggre-
gating a geographical information system (GIS)to a multiagent system in the simulation of so-
cial or urban environments characterizes
geosimulation (Benenson and Torrens 2004).Despite recent proposals on new models and
implementation of instructional layers in sim-ulators (Gibbons et al. 2001), few tools havebeen created specifically for geosimulation.
These applications involve particular features,
such as geographic features, which must be ex-ploited by intelligent tutorial systems in order
to enrich learning.
This article describes the ExpertCop educa-
tional tutorial system, considering the primarycharacteristics we claim to be essential to the
Articles
FALL 2006 63Copyright 2006, American Association for Artificial Intelligence. All rights reserved. ISSN 0738-4602
A MultiagentSimulator for
Teaching PoliceAllocation
Vasco Furtado and Eurico Vasconcelos
A I M agaz in e V o l u m e 2 7 N u m b e r 3 ( 2 0 0 6 ) ( A A A I )
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
2/12
the activity studied allows the user to learn bydoing (Piaget 1976) and to understand thecause-and-effect relationship of his or her ac-tions. According to Kolb (1984), learning is fa-vored when the learning process occurs withinthe following four successive steps:
Concrete Experience: Obtained through the ac-tivity itself or its simulation in a virtual envi-
ronment.Reflexive Observation: The experience is fol-lowed by the reflection phase. It is recreated in-ternally in the users mind under different per-spectives.
Abstract Conceptualization: In this stage, theexperience is compared and its patterns,processes, and meanings are analyzed. Withinthis context, abstract concepts and new knowl-edge are created. The knowledge is generated intwo moments of the cycle, in this step and inthat of concrete experience. The knowledgegenerated in the concrete experience phasecomes only from the simple observation of the
external event, while the knowledge generatedin the abstract conceptualization phaseemerges as a consequence of an internal cogni-tive process of the student.
Active Experimentation: In this stage the stu-dent will conduct a new experiment with thenewly acquired or modified concepts.
The simulation per se is not a sufficient toolfor education. It lacks the conceptual ability onthe part of the student to understand the sim-ulation model. Therefore, some works (Taylorand Siemer 1996), (Angelides and Siemer, 1995)have tried to integrate the notions of intelli-
gent tutoring system (ITS) and simulation inorder to better guide learning and to improveunderstanding of the simulation process. Theidea of an ITS is the integration of artificial in-telligence in computer learning systems. Itaims at emulating the work of a human teacherwho has knowledge of the content to betaught, as well as how and to whom it shouldbe taught. To achieve this, we need to represent(1) the domain of study, (2) the pedagogicalstrategies, and (3) the student to whom theteaching is provided. A fourth component mayalso be considered (Kaplan and Rock 1995,Woolf and Hall 1995)the interface with the
user. The user interface determines how the in-teraction with the system is. Through the in-teraction of these components, the ITS adaptspedagogical strategies on a domain at the levelof the student for his or her individual needs.
The ExpertCop System
Police resource allocation in urban areas to per-form preventive policing is one of most impor-
general architecture of an educational geosimu-lation. ExpertCop aims to enable police officersto better allocate the preventive police force inurban areas. This software produces, based on apolice resource allocation plan, simulations ofhow crime behaves in a certain period of timebased on the defined allocation. The goal is toallow a critical analysis by police officers who
use the system, making them understand thecause-and-effect relation of their decisions.Geosimulation generates a great amount of
data deriving from the interactions occurringin the simulation process, and it is necessary tomake chronological, geographical, and statisti-cal associations among these data to under-stand the cause and effect of the simulatedevents. Thus, we propose the use of an intelli-gent tutor agentthe pedagogical agentas adata analysis supporting tool. This agent uses amachine-learning concept formation algo-rithm to identify patterns on simulation data,to create concepts representing these patterns,
and to elaborate questions to the student aboutthe concepts learned. Moreover, it explores thereasoning process of the domain agents by pro-viding explanations, which help the student tounderstand simulation events.
Urban Simulationand Intelligent Tutoring
Simulations based on multiagent systems arelive simulations that differ from other types ofcomputer simulations because simulated enti-ties are individually modeled through the use
of agents. According to Gilbert (Gilbert andConte 1995), the multiagent approach (bot-tom-up) is appropriate for the study of socialand urban systems. Social or urban environ-ments are dynamic and nonlinear and are com-posed of a great number of variables. Multia-gent systems are also appropriate when theenvironments are made up of a great numberof entities whose individual behaviors are rele-vant in the general context of the simulation.
A particular kind of simulation, calledgeosimulation, addresses an urban phenomenasimulation model with a multiagent approachto simulate discrete, dynamic, and event-ori-ented systems (Benenson and Torrens 2004). Ingeosimulated models, simulated urban phe-nomena are considered a result of the collectivedynamic interaction among animate and inan-imate entities that compose the environment.The geographic information system (GIS) is re-sponsible for providing the data ware ingeosimulations.
Simulation is widely used as an educationaltool because the computerized simulation of
Articles
64 AI MAGAZINE
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
3/12
tant tactical management activities and is usu-ally decentralized by subsectors in policeprecincts of the area. Tactical managers analyzethe disposition of crime in their region and ac-cordingly allocate the police force. We adopt theprinciple that by knowing where crime is hap-
pening and the reasons associated to such crime,it is possible to make an optimized allocationand, consequently, to decrease the crime rate.
The volume of information that police de-partments have to analyze is one of the mainfactors to provide society with efficient an-swers. Tactical managers who perform police al-locations, for instance, lack ability for decisionmaking based on information analysis. In real-ity, understanding criminal mapping activities,even using GIS, is a nontrivial task. In additionto that, experiments in this domain cannot beperformed without high risks because they re-sult in loss of human lives. In this context, sim-
ulation systems for teaching and decision sup-port provide a fundamental tool.
Goals
The ExpertCop system aims to support educa-tion through the induction of reflection onsimulated phenomena of crime in an urbanarea. The system receives as input a police re-source allocation plan, and it makes simula-
tions of how the crime rate would behave in acertain period of time. The goal is to lead thestudent to understand the consequences of theallocation as well as understanding the cause-and-effect relations.
In the ExpertCop system, the simulations oc-
cur in a learning environment along withgraphical visualizations that help the studentslearning. The system allows the student to ma-nipulate parameters dynamically and analyzethe results.
ExpertCop Architecture
ExpertCop architecture is composed of a multi-agent system, a GIS, a database, and an inter-face as shown in figure 1. The interface in Ex-pertCop allows the student to move among thefunctionalities and processes of the system in alogical, ergonomic, and organized way. The GISis responsible for generating, manipulating, andupdating a map of the city to be studied in asmall scale. The system database contains (1)the information about each student and abouthis or her simulations, (2) the configuration da-ta, (3) the real data on crime and statistics oncrime yielded for the department of state police,and (4) the domain ontology. The most impor-tant component is the multiagent systems plat-form, and it will be discussed in detail next.
Articles
FALL 2006 65
SystemDatabase
Interface
ExpertCop Architecture
Interact
Multiagent Plarform JADE
Control Agents Pedagogical Agent
Domain Agents
Miner
Police Teams
Targets
Criminals
GIS
Interact
Figure 1. The ExpertCop System Architecture.
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
4/12
nected points. The patrol areas are given to thepolice team as a mission. These areas are asso-ciated to intervals of time so as to fill out thework period of the team. A team with a workshift of 8 x 16 should patrol an area (or areas)for repetitive periods of 8 hours every 24 hours.Suppose there is a police team called Beta whohas two areas to patrol in its shift. The user or
student should define the intervals of each areaas follows.
Beta police team:Work shift: 8 x 16; Work period: 8 AM to 4 PM.
Area of patrol 1(4hs): start 8 AM end 12 PM.
Area of patrol 2(4hs): start 12 PM end 4 PM.
The total work load is eight hours.
Each team organizes its patrol areas in a listaccording to the chronological order of thework period associated to each area. One ormore points that determine the area make upeach patrol area in its turn. Initially, the policeteams begin their activities at a common initialgeographical point (the neighborhood police
station). From that point onwards, the workingpolice teams (during their work period) verifythe schedule by which their areas must be pa-trolled. After identifying the area that must bepatrolled at that time, the police team places,in order, the points that form that area in a listand utilizes the first point as the objectivepoint. With the objective defined, the teamshould move towards it. To obtain the next po-sition at each moment of the simulation, theteam asks the GIS agent for the next point be-tween the current position and the objectivepoint according to the speed of the manner of
locomotion used.The calculation of walking time of the agenttakes into account the time elapsed betweenthe last point request and the current time.When arriving at its objective point, the teamplaces it at the end of the list and considers thenew initial point of the list as its objective. Fol-lowing this flow, the team moves along thepoints that make up the patrol area. Thisprocess of going to different patrol areas anddifferent patrol points is repeated until the endof the teams work shift.
CriminalsThe criminal manager creates each criminalagent in the simulation, with the mission ofcommitting a specific crime. After selection ofthe area and simulation period by the student,the criminal manager loads, from the systemdatabase, all the crimes pertaining to the areaand period selected and places the crimes inchronological order. When beginning the sim-ulation, observing the chronological order of
The Multiagent SystemsPlatform
The structure, communication, administration,and distribution of the agents is provided bythe Java Agent Development Framework
JADE.1 The multiagent platform in ExpertCopis made up of three groups of agents: control
agents, domain agents, and the pedagogicalagent. The control agents are responsible forthe control, communication, and flow in thesystem. The most important control agents arethe GIS agent, which is responsible for answer-ing requests from the graphical interface andthe domain and control agents; the manageragents, which are responsible for the coordina-tion and interaction with domain agents andcontrol preprogrammed activities such as acti-vation and deactivation; and the log agent,which is responsible for recording all interac-tions among system agents. Another importantcontrol agent is the pedagogical agent (PA). It is
endowed with pedagogical strategies, and aimsto help the user in understanding the simula-tion process and results. The PA will be dis-cussed in detail in the pedagogical proposalsection of this work.
The domain agents are the actors of the do-main. They are notable points, police teams,and criminals.
Notable Points
Notable points are buildings relevant to the ob-jective of our simulation, such as shopping
centers, banks, parks, and drugstores. They arelocated on the simulation map and have thesame characteristics as the buildings they rep-resent.
Police Teams
The mission of the police teams is to patrol theareas selected by the student during the workperiod and work shifts scheduled for the team.A software agent represents each team and hasa group of characteristics defined by the stu-dentsuch as means of locomotion, type ofservice, and work shiftwhich will influencethe patrol. The team works based on its workperiod and work shift. The work period deter-mines the beginning and end of work, and thework shift determines the work and rest peri-ods. Consider a team that works a shift of 8hours and then rests for 16, working from 8 AMto 4 PM. The team would work a first shift from8 AM to 4 PM, and then rest for 16 hours, re-turning to work the following day at 8 AM. Thepatrol areas are composed of one or more con-
Articles
66 AI MAGAZINE
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
5/12
the events, it creates a criminal agent for eachcrime. The criminals task is to evaluate the vi-ability of committing the crime. The evalua-tion is based on risk, benefit, and personalityfactors, defined on the bases of a set of inter-views with specialists on crime from the PublicSafety Secretariat and on research in the area ofcriminal psychology.
The risk is defined starting from the vari-ables.
Type of crime: Each type of crime is associat-ed with a risk level, which is based on the typeof punishment for the crime, on the level of ex-perience, and on the apparatus of the criminal.ExpertCop works only with robberies, thefts,and burglaries, which are types of crime influ-enced directly by preventive policing.
Type of target: The type of target indicates theresistance capacity against a crime. These tar-gets are associated with the types of crime men-tioned previously. Table 1 associates the riskvalue of the type of crime with the type of tar-
get.Police presence: Police presence (distance in
relation to the place of the crime) is the mainfactor that influences risk. The greater the dis-tance between the closest team and the placeof the crime, the lower the risk is. We consid-ered three categories for the evaluation of thecriminal as to the distance from policing. Anydistance between 0 and 200 meters is consid-ered close, between 200 and 500 meters is con-sidered as average distance, and above 500 me-ters is considered far. These categories weredefined by an experienced police officer basedon the average persons visual range and the av-erage length of a city block.
Public illumination: When the crime occurs atnight, public illumination in the area is a factorof evaluation. Areas with deficient illuminationfacilitate criminal action and directly influencethe risk. The areas can be classified as poorly orwell illuminated.
Existence of escape routes: The existence ofplaces such as slums, woods, or deserted areasclose to the place of the crime facilitates escape,
augmenting the risk that the crime will be com-mitted. The classification as to the proximity ofescape routes follows the same parameters asthe distances of police teams. These areas maybe close (0 to 200 meters), at average distance(200 to 500 meters), or far away (above 500 me-ters) from the place of the crime.
Benefit is defined starting at the type and
amount of spoils that the target can offer. Forexample, a person is low while a bank is veryhigh.
Personalitydefines criminal courage levelin the face of crime. When being created, a typeof personality is associated to the criminal (ap-prehensive, careful, bold, and fearless) chosenrandomly by the criminal manager and givingrandom airs to the criminal. A bold criminalevaluates risk with fewer criteria, giving moreweight to the benefit. But an apprehensivecriminal does the opposite, giving much moreweight to the risk.
The values of the variables regarding crime(type of crime, type of target, geographical lo-cation of crime, date, and time) are sent to thecriminal by the criminal manager. But to ob-tain the data on the environment (geographi-cal factors), the criminal exchanges messageswith the GIS agent, which furnishes the geo-graphical location, date, and time of the crime.
Having collected all the necessary informa-tion for the decision support process of thecrime to be executed, the agent uses a set ofproduction rules for evaluating the viability ofcommitting the crime. The inference rules con-taining the structure of the decision support
process and an inference machine is represent-ed in the JAVA-based JEOPS shell (Figueira andRamalho 2000). This process results in the de-cision of whether or not to commit the crime.Figure 2 contains an example of rules. After de-ciding whether to commit a crime or not, thecriminal sends a message to the GIS agent,which then marks the decision on the map dis-played to the user (red if the crime is commit-ted; green if it is not).
Articles
FALL 2006 67
IF distance_police = close AND type_crime = robberyAND type_victim = bank THEN risk = highIF type_victim = bank THEN benefit = highIF benefit = high AND risk = high AND personality = boldTHEN decision = commit_crime
Figure 2. Example of Rules.
Capital letters denote the logical structure of the rule; bold-face letters represent the variables that make up the
agents internal state; italic letters represent the values of the variables coming from the data of the crime and the
exchange of messages with the GIS agent.
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
6/12
on the map is represented by a point in the col-or green (crime prevented) or red (crime occur-ring); these points function as hyperlinks to theexplanation tree of the agents decision-makingprocess. When the student clicks a point, thepedagogical agent will present an explanationcontaining the crime data and a hyperlink tothe explanation of the reasons this crime wascommitted or prevented. Follow-up questionscan be done for the comprehension of the con-cepts of the domain. This type of explanation isobtained in ExpertCop because the agents ar-chitecture has a cognitive module modeled ex-plicitly in terms of ontologies, production rules,and problem-solving methods (see Furtado and
Vasconcelos [2007] for details of this modeling).The tree of proofs is represented in node sets ofthe proof markup language (PML) (Pinheiro daSilva, McGuinness, and Fikes 2006) which is oneof the components of the inference web infra-structure. Thus, proof fragments in PML can beshared with other applications, besides usingthe inference web infrastructure to abstractproofs into explanations and to present proofsand explanations to users.
Macrolevel Explanation
In ExpertCop, we understand as emerging be-havior the effects of individual events in crime,its increase or reduction, criminal tendencies,and seasonableness. For the explanation of theemerging behavior of the system, the pedagog-ical agent tries to identify patterns of behaviorfrom the database generated in the simulation.
First, the agent takes (requesting the logagent) the simulation data (events generated forthe interaction of the agents as crimes [date,hour, motive, type] and patrols [start time, final
The Pedagogical Proposal
of the SystemThe pedagogical model of the system is basedon the concept of the intelligent tutoring sim-ulation system, which includes the simulationplus an agent that provides adaptive explana-tions for a student.
The Simulation as aPedagogical Tool
ExpertCop simulation is designed to be part ofa pedagogical tool. The student can learn by
doing. He or she initially interacts with the sys-tem allocating the police, which exposes his orher beliefs about the allocation of resources. Asimulation of the agents interaction is thendone and the student beliefs can be validatedby means of a phase of result analyses. This cy-cle can be repeated as many times as the stu-dent finds necessary.
The pedagogic agent uses two distinct formsto explain the events of the system, the expla-nation at a microlevel and at a macrolevel.
Microlevel Explanation
The microlevel explains the simulation events(crimes). ExpertCop uses a tree of proofs de-scribing the steps of reasoning of the criminalagent responsible for the event. This tree is gen-erated from the process of the agents decisionmaking stored in the database. The student canobtain the information on the crime and theprocess that led the agent to commit it or not byjust clicking the point that represents the crimeon the map with the mouse. Each crime located
Articles
68 AI MAGAZINE
Target/Crime Robbery Theft Burglary
Person Low Very Low x
Vehicle Average Very Low x
Drugstore Average Very Low x
Lottery House High X x
Gas Station Average X x
CommercialEstablishment High Low Low
Bank Very High X x
Residence Average Low Low
Table 1. Table of Risks per Type of Crime and Target.
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
7/12
time, stretch]) and preprocesses it, adding geo-
graphic information such as escape routes; co-ordinates of notable places; distance between
events, agents, and notable places; and social
and economical data associated to geographic
areas. After preprocessing, in the mining phase,the PA identifies patterns by means of the prob-
abilistic concept formation algorithm COBWEB
(Fisher 1987), which generates a hierarchy ofprobabilistic concepts. Probabilistic concepts
have attributes and values with an associated
conditional probability of an entity having an
attribute a with a value vgiven the fact that thisentity is covered by the concept C, P(a = v|C).
The generated concepts are characterized ac-
cording to their attribute or value conditionalprobabilities. That is to say, a conceptual de-
scription is made of attribute or values with
high probability. Having the probabilistic con-
cept formation hierarchy constructed, the agentidentifies and filters the adequate concepts for
being transformed into questions to the stu-
dent. The heuristics used to filter which con-cepts will generate questions to the student and
which features will compose these questions
follow the steps below. The root of the hierar-
chy is ignored (not appraised), because it aggre-gates all the concepts and is thus too general.
The hierarchy is read in a bottom-up fashion
from the most specific to the most generic con-
cepts. The criteria used in the analysis of theconcepts for selection are (1) concept must cov-
er at least 10 percent of the total of exam-pleswe assume that fewer than 10 percent of
the examples would make the concept poorlyrepresentative; (2) an attribute value is exhibit-ed in the question only when it is present in at
least 70 percent of the total of the observationscovered by an example; and (3) a question must
contain at least three attributes.When going through a branch of the tree
considering the previous items, in case a con-
cept is evaluated and selected, the nodes supe-rior to this concept (parent, grandparent ...)will no longer be appraised to avoid redundant
information. This doesnt exclude the nodes inthe same level of the hierarchy of this node
that may be appraised in the future.An example of the Cobweb result is the con-
cept depicted in figure 3. That concept is dis-
played to the user or student as the followingquestion: Did you realize that crime: theft,victim: vehicle, day of the week: Saturday, pe-
riod: night, place: residential street, neighbor-
hood: Aldeota frequently occur together?Having this kind of information, the user or
student can reflect on changes in the alloca-tion, aiming to avoid this situation.
System Function
Initially, the student must register with the sys-
tem and configure the simulation parametersusing a specific interface. After that, the stu-
Articles
FALL 2006 69
Day of Week [Saturday] 100%Crime type [Theft] 95%
Crime type [ Robbery ] 5%Object [Vehicle] 100%Police team [Far from] 60%Police team [Near] 40%Neighborhood [Aldeota] 90%Neighborhood [Papicu] 10%In traffic light [Yes] 45%In traffic light [ No] 55%Period [Night] 100%Place [Residential street] 100%
concept
concept_1
concept_2
C_1.1
C_1.1.1
C_1.1.2
C_1.2C_1.3
C_1.4
C_2.1
C_2.2
C_2.3
C_1.1.3
Figure 3. Example of ExpertCopy Concept Tree Formed by PA.
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
8/12
(described previously). A microlevel explana-tion can be obtained by clicking any red or
green points on the screen, which indicatecrimes that occurred or that were prevented, re-spectively. The student can request a macrolev-
el explanation by pressing the hint button rep-resented on the screen. A set of questions is
shown to the student in order to make him orher reflect about poossible patterns of crimes.
At each new allocation performed, the sys-
tem will comparatively evaluate the simulatedmoments, showing the student whether or notthe modification brought about a better effect
upon the crime rate. PA also makes compar-isons among results obtained in each simula-
tion tour to evaluate the students learning im-provements. The student can also evaluate theresults among a series of simulations on the
evaluation screen. On this screen, the results ofall simulations made by the student are shownin a bar graph.
Evaluation of the System
ExpertCop was used to support a course at theBrazilian Ministry of Justice and the NationalSecretariat of Public SafetySENASP. The ob-
jective of this course was to emphasize the im-portance of information technologies in publicsafety. ExpertCop was intended to help police
officers reflect on the forms of treatment andanalysis of information and how this influ-
dent determines the number of police teams to
be allocated and the characteristics of theseteams. Based on the geographic and statistical
data available in the map about the area and on
his or her knowledge about police patrol, the
student determines the areas to be patrolledand allocates the police teams available on the
geoprocessed map. To realize the allocation
process, the student selects the patrol areas inthe map for each team. After that, he or she de-
fines the period of time that the police team
will be in each patrol area. The sum of each pe-
riod of time must be equal to the teams work-load. Figure 4 shows the interface for the allo-
cation process.
Agents representing the police teams moni-tor the patrol areas defined by the user follow-
ing the programmed schedule. The patrol func-
tion is to inhibit possible crimes that could
happen in the neighborhood. We presume thatthe police presence is able to inhibit crimes in
a certain area size. The goal of the student is to
provide a good allocation that prevents thehighest number of crimes.
After the configuration and allocation
process, the user can follow the simulation
process in the simulation interface. At the endof the simulation process, the user accesses the
pedagogical tools of the system. Figure 5 shows
the functionalities for visualization.Besides the visualization functionalities, the
student can access the explanation capabilities
Articles
70 AI MAGAZINE
Figure 4. ExpertCops Police Allocation Interface.
}Fields for selecting the police team and
workload.
} Space for association of areas to bepatrolled with periods of time. Each areacorresponds to a different color.
Patrol areas created by the user.
Buttons for map manipulation (clean map,zoom, creation of areas, undo).
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
9/12
ences the understanding of crime. The audi-ence was made up of 30 professionals in the
area of public safety: civil police officers, chiefsof police, and military police (which are themajority). These groups of professionals com-
pose the target public towards which this toolis geared.
MethodologyExpertCops workflow tries to improve the
learning process proposed by Kolb (1984) inthe following four successive steps: (1) the
process of simulation as a concrete experience;(2) reflection and observation of the resultswith the aid of the support tools offered by the
system; (3) abstraction and conceptualization,supported by the hints offered by the pedagog-ical agent on the patterns revealed in the sim-
ulation; and (4) new experimentation with theconcepts acquired in a new process of simula-
tion. The use of ExpertCop occurred in two dis-
tinct stages, one explanatory and the otherevaluative.
In the first stage, the participants were in-structed on the process of allocating of policeresourceswhat it is all about, how it occurs in
practice, and the factors involved in thisprocess. After this contextualization, Expert-Cop was presented, with its objectives and
functionalities. After concluding this stage, theparticipants made use of the tool in an illustra-
tive simulation to familiarize themselves with
the functionalities.In the second stage, training was carried out
by a set of at least three simulations in city ar-
eas. In the first simulation, the participants had
to create and configure a certain number ofteams (according to the size of the area), allo-
cate them on the map, and activate the simu-
lation. At the end of the first simulation weasked the participants to identify, according to
their beliefs, five factors that influenced the oc-
currence of the crimes. They did so by observ-
ing the map of the crimes that occurred andthose that were prevented. We requested that
the participants not mention complex factors
of political or socioeconomic order, such as un-employment or taxes, because we focused on
geographical or visual factors that directly af-
fect the crime rates. After collecting the partic-
ipants beliefs, we allowed them to use the ped-agogical support of the system (clues,
explanations, and evaluations). After the use of
the pedagogical support tools, the collection ofbeliefs about the crimes was carried out again.In the subsequent simulation, we repeated the
same area to serve as a comparison with the ini-
tial simulation already completed, and allowedstudents to make their allocations and use the
pedagogical support of the tool according to
their needs. Afterwards, we performed two oth-
er simulations with areas different than the firstone. Using different areas for each simulation
Articles
FALL 2006 71
Figure 5. Visualization Functionalities.
}}
Map with the simulation results(crimes occurred and avoided)filtered by the user selections.
Hint button.
Functionalities for selection of thetype of graphic and the crimes byperiod (month, day, time).
Graphic about the selected results.
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
10/12
Based on an analysis of the relation betweenthe number of times the systems pedagogicalsupport was accessed and the results of the sim-ulation, it was possible to observe that studentswith a higher number of accesses obtained bet-ter results. We also observed that the average ofthe results of the simulations after the use ofthe pedagogical support (second and third sim-
ulations) was greater than the average of the re-sults in the simulation before the use of thesupport offered by the system. In further ex-periments with a set of 90 students, we mea-sured a statistically significant increase of 9.09percent. A detailed description of these andother results can be seen in Furtado and Vas-concelos (2007).
We also observed that although the studentsdemonstrate more ability in handling the toolas time goes by, they spend more time duringthe allocation process. We think that this indi-cates a more reflexive allocation process. An-other observation is that an atmosphere of co-
operation (or competition) was created amongthe students, and they often compared resultsand patrol routes seeking to identify similarstrategies among themselves. Effectiveness oflearning depends on the student profile. Anovice in computer science tends to concen-trate on the tool instead of on the allocationprocess. Experienced students in the allocationprocess improve their performance to a lesserdegree. Based on the results, we may also con-clude that the learning level is higher in partic-ipants with little or no experience in the do-main or in the treatment of information. Also
evaluating these results, we conclude that thepedagogical support offered by the systemhelps the participants understand and betteridentify the factors that affect crime, allowingthus for better performance in their allocationsand consequently a reduction in crime levels.The students were capable of noticing the im-portance of analyzing the data in the allocationprocess. The tool was revealed as easy to useand attractive to the students. They continueusing the system even after the end of thecourse.
We observed as a negative fact that two stu-dents who obtained very low results were not
motivated to follow the subsequent simula-tions. Another important aspect that merits dis-cussion is how learning is influenced by thequality of the model. In our discussions witheducators, we were advised to be careful thatstudents do not come away thinking that every-thing they did in ExpertCop will be reproducedequally in their area of work. For this reason, weprefer not to present students with the areaswhere they actually work. The objective of the
allows us to evaluate whether the student wasable to abstract the modified or acquired con-cepts by applying them in different contexts(characteristics), since the new concepts are ap-plicable to all of the characteristics that the en-vironment presents. During the simulations,the time needed to accomplish the allocationprocess in the training simulations was mea-
sured for a sample of the participants (two fromeach different group, civil police officers, chiefsof police, and military police).
We formulated the hypothesis that the sys-tem will make the students improve their un-derstanding of the results (causes and effects),acquiring new beliefs or modifying old beliefsregarding crime and the allocation process.Based on the belief that the percentile of crimesprevented in relation to the total number ofcrimes attempted represents the participantsperformance in a simulation, we formulatedthe second hypothesis: The acquisition of newconcepts and beliefs will make the students im-
prove their allocation and consequently obtainbetter results in the simulations.
Results and Discussion
From the analysis of the collected beliefs, wemade a number of observations. First, 87 per-cent of the students changed or included newbeliefs about motives and causes of crime. Sec-ond, in the evaluation, beliefs that were morespecific and practical replaced those initiallyobserved, which were more generic. Third, newbeliefs about specific factors such as public illu-
mination, patrol route distances, the existenceof slums, and work shifts, were included by thestudents in the second collection. Fourth, timefactors, such as the relationship between theday and the periods of the day with the num-ber of crimes occurred, began to be taken intoconsideration. Fifth, a large number of beliefswere mentioned related to the importance ofthe analysis of the characteristics of the geo-graphical area for good policing. Sixth, the mil-itary police that worked in the allocationprocess indicated a low alteration in their be-liefs, mentioning relevant factors at once in thefirst collection. They included factors that were
until then unknown to the system. We consid-er this to be important because it enables im-proving the system, and because we see thatthe goal of reflection on factors that influencethe crime was obtained even in this situation.
According to the data, it is possible to con-clude that our hypothesis is valid, confirmingthat the system offered subsidies so that thestudent modified his or her beliefs on the do-main or acquired new beliefs.
Articles
72 AI MAGAZINE
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
11/12
tool is to make them reflect on causesand effects considering that the stu-dent is capable, through practical situ-ations, of reflecting upon alternativesregarding allocation of police officers.
The matter of modeling the crimi-nal also deserves mention. Cognitivemodeling of the criminal is a difficult
task, and it is practically impossible toacquire and represent all the nuancesof its cognitive model. What is in-tended is to model the basic knowl-edge that is found in the literature aswell as heuristics supplied by police of-ficers. In this way, the explanation ofwhy crimes are committed leads thestudent to know some of the factorsthat were considered during the elab-oration of the crime. This knowledgeis shown opportunistically within aprecise context. If, for any reason (per-haps due to his or her field experi-
ence), the student is led to disagreewith the performance rules of thecriminal agent or to elaborate othercharacteristics that were not observed,ExpertCop has nonetheless fulfilled itsrole of making the student reflect up-on these factors. Moreover, maintain-ing these rules that determine the cog-nitive model of the criminals, sincethey are explicit in an ontology, can beeasily modified and adapted to differ-ent realities with no need to changethe source code.
Related Work
Previous use of multiagent systemssimulation in education (Khuwaja,Desmarais, and Cheng 1996; Querrecet al. 2004; Gibbons et al. 2001), ITS(Johnson, Rickel, and Lester 2000), so-cial simulation to support decisionmaking, and GIS tools (Gimblett 2002)strongly influenced this research work.Our proposal is an intersection amongthese areas. There are many projectsthat describe solutions with parts of
our system design. Virtual environ-ments for training, such as Secureviproposed by Querec (2004), is a systembased on the Mascaret model that us-es multiagent systems to simulate real-istic, collaborative, and adaptive envi-ronments for training simulation.Intelligent GIS, such as the systemproposed by Djordjevic-Kajan et al.(1995), intends to provide computer
support in fire rescue. The system hasa Fire Trainer, an intelligent agentthat covers the activities connected toeducation. The Phoenix system (Co-hen et al. 1989) is a discrete event sim-ulator based on an agent architecture.The system is a real time, adaptiveplanner that simulates the problem of
forest fires.A number of intelligent tutoring sys-tems have been designed, such as theone built by Wisher et al. (2001), whichdescribes an intelligent tutoring forfield artillery training, and the Sherlocksystem by Lesgold et al. (1992), whichprovides advice for impasses while us-ing a simulated system. The architec-ture proposed by Atolagbe and Hlupic(1996) and Draman (1991) for educa-tional simulation is similar to this work,although Atolagbe and Hlupic do notemphasize the power of simulation in
GIS or the use of knowledge discoveryand data mining (KDD) to improve stu-dent learning. Several works in gamesand entertainment (Galvo, Martins,and Gomes 2000; Leemkuil et al. 2003)use simulation with an educational pro-pose. Even though they present somesimilarities with our approach, gamesimulators have a different pedagogicalstrategy. They focus on the results ofthe simulation while we believe thatthe most important aspect is theprocess itself. Another differential isthat few games are adapted to the stu-dent level. In order to diminish this, wehave proposed putting ITS features ingames as was described by Angelidesand Siemer (1995).
Conclusionand Future Work
This article described the ExpertCopsystem, a pedagogical geosimulator ofcrime in urban areas. The ExpertCoparchitecture is based on the existenceof multiagent systems with a GIS toperform geosimulations and of a ped-agogical agent that follows the simula-tion process; the agent can definelearning strategies as well as use a con-ceptual clustering algorithm to searchfor relations in the facts generated inthe simulation. ExpertCop is focusedon police officers education, relatedto resources allocation.
Initial training sessions with police
officers interacting with the systemwere performed aiming to evaluatelearning by using this tool. As a com-plement to the use of the system, acourse was held where ExpertCop wasused as a tool for analysis and reflec-tion of practical situations. Themethodology adopted to analyze the
learning of students in ExpertCop hasshown an improvement in the stu-dents data analysis abilities, in theprocess of resource allocation with Ex-pertCop, and in the identification offactors that influence the crime.
We intend to continue the researchon the ExpertCop system, enhancingits functionalities such as rendering itmultiuser and making it available onthe web. Other ongoing work aims attransforming it into a decision-mak-ing support tool. To accomplish this,we are adopting a different approach
for the crime simulation model. Thecriminal model is based on self-orga-nization systems inspired by biologi-cal systems. Self-organization refers tothe phenomenon of a society ofagents that demonstrate intelligentbehavior (as a collective) out of simplerules at the individual level. We mod-el the criminals as distributed entitiesby agents with the ability to demon-strate self-organization from their in-dividual (local) activities as well as tak-ing into consideration the influence of
other criminals in the community inwhich they live (Furtado et al. 2006).We are also designing an evolutionaryapproach that integrates with the sim-ulation tool and is devised to assist po-lice officers in the design of effectivepolice patrol routes. Our approach isinspired by the increasing trend of hy-bridizing multiagent systems withevolutionary algorithms. Our idea is touncover strategies for police patrollingthat cope with the dynamics of thecrime represented by criminals thatlearn on the fly. To uncover good po-
lice patrol routes in this context, weare integrating into the simulationmodel a genetic algorithm. Prelimi-nary experiences have shown thatsuch an approach is very promising(Reis et. al. 2006).
Acknowledgement
We thank Bill Clancey for his insight-ful comments.
Articles
FALL 2006 73
-
7/29/2019 A Multi-Agent Simulator for Teaching Police Allocation
12/12
Wisher, R. A.; MacPherson, D. H.; Abram-
son, L. J.; Thorndon, D. M.; Dees, J. J. 2001.The Virtual Sand Table: Intelligent Tutoring
for Field Artillery Training. Alexandria, VA:
U.S. Army Research Institute for the Behav-
ioral and Social Sciences.
Woolf, B.; and Hall, W. 1995. Multimedia
Pedagogues Interactive Systems for
Teaching and Learning, IEEE Computer
28(5): 7480.
Vasco Furtado is profes-sor of computer science
at University of Fortaleza
(UNIFOR), Brazil, where
he also leads a team of
researchers in theKnowledge Engineering
group that studies agent-
based simulation and
agents explanation on the web. He is con-
sultant of the State of Ceara IT Company(ETICE) where he has coordinated and de-
veloped research and development projectson the law enforcement domain. Furtado
holds a Ph.D. in Informatique from the
University of Aix-Marseille III, France. He ison sabbatical in the Knowledge Systems
Laboratory at Stanford University in
200607. Further information about publi-
cations and projects is available at
http://www.mentores.com.br/ vasco/in-dex.html.
Eurico Vasconcelos re-
ceived a Master of Sci-
ence degree in computer
science from the UNI-FOR and is currently
pursuing a Ph.D. at the
Catholic University of
Rio de Janeiro, Brazil.
His interests include in-telligent tutoring systems, human-comput-
er interaction, and multiagents. He can be
reached at [email protected].
Gilbert, N.; and Conte, R., eds. 1995.Artifi-cial Societies: The Comuter Simulation of So-cial Life. London: UCL Press.
Gibbons, A. S.; Lawless, K. A.; Anderson, T.
A.; and Duffin, J. 2001. The Web and Mod-
el-Centered Instruction. In Web-basedTraining, ed B. H. Khan, 137146. Engle-wood Cliffs, NJ: Prentice Hall.
Gimblett, H. R., ed. 2002. Integrating Geo-
graphic Information Systems and Agent-BasedModeling Techniques for Simulating Social andEcological Processes. London: Oxford Uni-versity Press.
Johnson, W. L.; Rickel, J. W.; Lester, J. C.
2000. Animated Pedagogical Agents: Face-
to-Face Interaction in Interactive Learning
Environments.International Journal of AI inEducation 11: 4778.
Kaplan, R.; and Rock, D. 1995. New Direc-
tions for Intelligent Tutoring Systems.AIExpert10(1): 30-40.
Khuwaja, R.; Desmarais, M.; Cheng, R.
1996. Intelligent Guide: Combining User
Knowledge Assessment with PedagogicalGuidance. In Proceedings of InternationalConference on Intelligent Tutoring Systems.Berlin: Springer Verlag.
Kolb, D. A. 1984.Experiential Learning: Ex-perience as the Source of Learning and Devel-opment. Englewood Cliffs, NJ: Prentice Hall.
Leemkuil, H. H.; de Jong, T.; de Hoog, R.;
and Christoph, N. 2003. KM Quest: A Col-
laborative Internet-Based Simulation
Game. Simulation & Gaming34(1): 89111.
Lesgold, S.; Lajoie, M.; Bunzo; Eggan, G.
1992. SHERLOCK: A Coached Practice En-
vironment for an Electronics Troubleshoot-
ing Job. In Computer Assisted Instruction andIntell igent Tutoring Systems. Hillsdale, NJ:Lawrence Erlbaum Associates.
Piaget, J. 1976.Le Comportement, Moteur delvolution. Paris: Ed. Gallimard.
Pinheiro da Silva, P.; McGuinness D. L.; and
Fikes, R. 2006. A Proof Markup Language
for Semantic Web Services.Information Sys-tems, 31(4-5): 381395.
Querrec, R.; Buche C.; Maffre E.; Chevailli-
er, P. 2004. SecuReVi: Virtual Environments
for Fire Fighting Training. Paper presented
at the Conference Internationale sur la
Ralit Virtuelle, 169175, Laval, France.
Reis, D.; Melo, A.; Coelho, A. L.; and Furta-
do, V. 2006. GAPatrol: An EvolutionaryMultiagent Approach for the Automatic
Definition of Hotspots and Patrol Routes.
In Proceedings of the Brazilian Symposium
of Artificial Intelligence, Ribeiro Preto, SP,
Brazil. Berlin: Springer Verlag.
Taylor, S. J. E.; and Sierner, J. 1996. En-
hancing Simulation Education with Intelli-
gent Tutoring Systems. InProceedings of the28th Winter Simulation Conference, 675680.New York: ACM Press.
Note1. TILab S.p.A. Java Agent DevelopmentFramework. JADE. 2003. At sharon.cselt.it/
projects/jade.
ReferencesAngelides, M. C., Siemer J. 1995. Evaluating
Intelligent Tutoring with Gaming-Simula-
tions. InProceedings of 1995 Winter Simula-tion Conference, 13761383. New York: As-sociation for Computing Machinery.
Atolagbe, T., Hlupic, V. 1996. A Generic Ar-
chitecture for Intelligent Instruction for
Simulation Modelling. InProceedings of the1996 Winter Simulation Conference,856863, ed. J. M. Charnes, D. J. Morrice,
and D. T. Brunner. New York: Association
for Computing Machinery.
Benenson, I., and Torrens, P. M. 2004.
Geosimulation: Object-Based Modeling ofUrban Phenomena. Computers, Environmentand Urban Systems 28 (1/2): 18.
Cohen, P. R.; Michael L. G.; David M. H.;Adele E. H. 1989. Trial by Fire: Understand-
ing the Design Requirements for Agents inComplex Environments.AI Magazine10(3):3248.
Djordjevic-Kajan, S.; Mitrovic, D.; Mitrovic,
A.; Stoimenov, L.; Stanic, Z. 1995. Intelli-gent GIS for Fire Department Services. Pa-
per presented at Eurocarto XIII, 177185.
Ispra, Italy, October 24.
Draman, M. 1991. A Generic Architecture
for Intelligent Simulation Training Sys-tems. InProceedings of the 24th Annual Sym-posium on Simulation, 3038. Los Alamitos,CA: IEEE Computer Society.
Figueira, F. C., and Ramalho, G. 2000.JEOPSThe Java Embedded Object Produc-tion System. InAdvances in Artificial Intelli-gence, 5261. Berlin: Springer-Verlag.
Fisher, D. 1987. Knowledge Acquisition Via
Incremental Conceptual Clustering. Ma-chine Learning2(1):139172.
Furtado, V., Melo, A., Menezes, R., Belchior,
M. 2006. Using Self-Organization in an
Agent Framework to Model Criminal Activ-
ity in Response to Police Patrol Routes. InProceedings of the Florida Artificial IntelligenceResearch Society Conference.Menlo Park, CA:AAAI Press.
Furtado, V., and Vasconcelos, E. 2007.Geosimulation in Education: A System for
Teaching Police Resource Allocation.Inter-national Journal of Artificial Intelligence in Ed-ucation 17(1), forthcoming.
Galvo, J. R.; Martins, P. G.; Gomes, M. R.2000. Modeling Reality with Simulation
Games for a Cooperative Learning. In Pro-ceedings of the 2000 Winter Simulation Con-ference, 16921698. Orlando, FL: Society forComputer Simulation.
Articles
74 AI MAGAZINE
Please Join Us forAAAI-07
2226 July 2007
Vancouver
British Columbia,
Canada