You are on page 1of 13

Decision Support Systems 75 (2015) 2537

Contents lists available at ScienceDirect

Decision Support Systems

journal homepage: www.elsevier.com/locate/dss

A Decision Support System for predictive police patrolling


M. Camacho-Collados a,b, F. Liberatore c,
a
Spanish National Police Corps, Madrid, Spain
b
Statistics and Operations Research Department, University of Granada, Granada, Spain
c
AIDA, Computer Science Department, Autonomous University of Madrid, Madrid, Spain

a r t i c l e i n f o a b s t r a c t

Article history: In the current economic climate, many police agencies have reduced resources, especially personnel, with a
Received 2 September 2014 consequential increase in workload and deterioration in public safety. A Decision Support System (DSS) can
Received in revised form 1 March 2015 help to optimize effective use of the scarce human resources available. In this paper we present a DSS that merges
Accepted 24 April 2015
predictive policing capabilities with a patrolling districting model, for the design of predictive patrolling areas.
Available online 4 May 2015
The proposed DSS, developed in close collaboration with the Spanish National Police Corps (SNPC), denes
Keywords:
partitions of the territory under the jurisdiction of a district that are efcient and balanced at the same time,
Predictive policing according to the preferences of a decision maker. To analyze the crime records provided by the SNPC, a method-
Time series forecasting ology for the description of spatially and temporally indeterminate crime events has been developed. The DSS has
Police Districting Problem been tested with a case study in the Central District of Madrid. The results of the experiments show that the
Multi-criteria decision-making proposed DSS clearly outperforms the patrolling area denitions currently in use by the SNPC. To compare the
Decision Support Systems solutions in terms of efciency loss, we discuss how to build an operational envelope for the problem considered,
which can be used to identify the range of performances associated with different patrolling strategies.
2015 Elsevier B.V. All rights reserved.

1. Introduction aim of providing an analysis of the evolution of crime in a territory.


More recently, both academics and practitioners, such as the RAND cor-
In this paper, we propose a Decision Support System (DSS) for the poration and the National Institute of Justice of the United States (NIJ),
implementation of a new paradigm of predictive police patrolling for have recognized the need for taking a step forward and developing
the efcient distribution of police ofcers in a territory under the juris- explicit DSS to provide help to decision makers in law enforcement
diction of a police department, with the aim of reducing the likelihood agencies [38].
of criminal acts. This DSS, called the Predictive Police Patrolling DSS
(P3-DSS), has been developed in collaboration with the Spanish Nation- 1.1. The Predictive Police Patrolling DSS (P3-DSS)
al Police Corps (SNPC).
Intuition has always been a fundamental part of police work [39]. In Spain, the security of towns is the responsibility of the SNPC, usu-
Knowledge in the eld by experienced managers of public security has ally sharing a territory with other local security forces. The SNPC is an
proven to be a useful weapon against crime. In most police departments, armed institution of a civil nature, dependent on the Spanish Ministry
this experience-based system remains unchanged. However, police in- of Home Affairs. Among its duties are keeping and restoring order and
tuition may not be taking into account all the factors inuencing the public safety and preventing the commission of criminal acts. The
evolution of crime. Historically, criminology has shown great interest SNPC is one of the country's most valued institutions, and is at the global
in the identication of areas with a higher rate of criminal activity [5]. forefront of the ght against crime, with the aim of constant innovation.
It has been proven that studying historical data allows identifying places Under the current system, the distribution of agents is determined by
where crime tends to agglomerate. Therefore, making use of historical the inspectors that coordinate the service during a particular shift.
information is fundamental for decreasing crime, as we know that Their experience, accompanied by preliminary information, such as a
crime has greater chances of happening when there are no security summary of the criminal activity of the last days, leads them to decide
measures and a motivated criminal encounters an appropriate objective on the allocation of policemen in a whole district.
[48,49]. Consequently, in the last decade, predictive policing measures The socio-economic context in recent years in Spain is that of a se-
have been developed, with different levels of sophistication, with the vere crisis that has reduced the number of police ofcers available to
the SNPC. Therefore, designing the distribution of agents in a territory
Corresponding author.
has become a complex task, and the lack of personnel can result in a
E-mail addresses: mcamacho0007@policia.es (M. Camacho-Collados), lowered level of security and, as a consequence, in an increased level
federico.liberatore@urjc.es (F. Liberatore). of crime. In order to continue providing the same level of security to

http://dx.doi.org/10.1016/j.dss.2015.04.012
0167-9236/ 2015 Elsevier B.V. All rights reserved.
26 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

the citizens of Spanish cities, the SNPC is taking steps to increase its most profound way. CompStat combined Geographic Information System
competitiveness, such as the development of a DSS to assist decision- (GIS) and crime mapping techniques to identify areas of high crime inten-
making processes in matters of public security in Madrid, with the sity. The importance of measuring the occurrence of crimes in the police
intention of applying this methodology to other cities. The seminal districts and keeping track of the actions of the police managers for
work of [47] showed that making use of optimization models results decision-making was proven by Weisburd et al. [55]. This topic was
in a reduction of the level of subjectivity present in this kind of opened to different approaches by the International Journal of Forecast-
decision-making, an improvement in the quality of the decisions, and ing, which published a special issue on crime forecasting in 2003 [24].
an increase in the level of satisfaction of the policemen involved. Years later, researchers at UCLA started a new approach to the inves-
This article has made the following contributions to the literature on tigations of crime agglomerations, modeling the dynamics of crime
DSS for public security management. First, we illustrate the pilot study hotspots and determining the parameter values that lead to the creation
that was undertaken in collaboration with the SNPC to develop a DSS of stable hotspots [45]. In a subsequent study, they used amplitude
for the implementation of a predictive police patrolling paradigm: the equations to study the development of crime hotspot patterns [44]
P3-DSS. This DSS provides predictive policing capabilities for forecasting and self-exciting point processes [35]. Also, they mathematically proved
the distribution of crime risk in a territory, as well as an optimization that there were different types of hotspots, even though they seemed
system that exploits this information to distribute agents in the best similar at rst sight. This breakthrough was further developed using
possible way, according to the preferences of the decision maker. To Levy Flight models by Chaturapruek et al. [14]. More recently, Zipkin
the best of the authors' knowledge, this is the rst DSS for public secu- et al. [58] introduced a police behavior component aiming at suppress-
rity managers that combines predictive policing techniques to support ing hotspots of criminal activity. In this model, the police deployment
the allocation of human resources. Furthermore, we present a method- adapts dynamically to changing crime patterns, making criminals mod-
ology for the temporal and spatial description of crime events in which ify, to a certain degree, their awareness and their criminal actions.
the time and the location of the incidents are indeterminate. In fact, the Probably the most ambitious predictive policing project so far made
exact time and location of occurrence of an incident is often not known use of the algorithms created by Brantingham and Mohler, along with
to the victim or the police. As an example, a victim of a pickpocket will LAPD Captain Sean Malinowski. With three years of data, and focusing
often realize that he/she has been robbed after a certain time, therefore on three types of crime in particular (i.e., burglary, automobile theft,
making it impossible to determine precisely when and where the crime and theft from automobiles), the algorithm points out areas of likely
occurred. Previous works on the subject have dealt exclusively with crime incidence. The rst analyses have shown a reduction of property
temporally indeterminate crimes [30] and, to the best of the authors' offenses where this methodology has been implemented, reporting
knowledge, no treatment for spatial uncertainty has been proposed in considerable reductions in serious violence crimes in the treatment
the literature. Moreover, we propose a novel algorithm to dene the cities and areas relative to comparison cities and areas. Another experi-
operational envelope for the specic shift under study. The operational ment of predictive policing was implemented in Santa Cruz, where
envelope is a powerful tool for identifying the range of impacts of differ- predictive maps based on risk percentages were given to security man-
ent patrolling strategies and to quantify possible efciency losses of agers. The use of these maps resulted in a 19% drop in burglaries [21].
suboptimal plans. Finally, to test our DSS, we present a case study of Another line of research that has been widely applied in practice has
the Central District of Madrid. This same methodology can be easily ex- focused on Risk Terrain Modeling (RTM) [11]. According to its creators,
tended to any district. Additionally, we provide some insights into the RTM is an approach to risk assessment in which separate map layers
patrolling strategies produced by the DSS and compare their perfor- representing the inuence and intensity of a crime risk factor at every
mance to the current patrol sector congurations adopted by the SNPC. place throughout a geography is created in a Geographic Information
The reminder of the article is organized as follows. In the next sec- System (GIS). Then all map layers are combined to produce a composite
tion we present the current state of the art of predictive policing. This risk terrain map with values that account for all risk factors at every
part also examines the theoretical background and current research place throughout the geography [10]. RTM has also been proposed as
into the Police Districting Problem (PDP) and focuses on the existing a methodology for the identication of risk clusters and the distribution
DSS for efcient policing. In Section 3, the structure of the proposed of police resources [30].
DSS, called P3-DSS, is given. Next, in Section 4, we present the operation- A number of models making use of methodologies other than hotspot
al envelope, its application in the evaluation of patrolling congurations, and RTM have been presented in the academic literature. Xue and Brown
and an algorithm for its computation. In Section 5 we apply the P3-DSS [56] and Smith and Brown [46] developed a spatial choice model and rep-
to a real case study of the Central District of Madrid. The article resented criminal events as point processes combining discrete choice
concludes with a summary of the main ndings of this research and techniques and data mining. They used this approach to predict the spa-
some possible future lines of research to be explored. tial behavior of criminals, comparing it with existing hotspot analyses.
Furtado et al. [22] model criminal behavior by using ant-inspired systems,
2. Related work trying to discover strategies for efcient police patrolling that take into ac-
count the dynamics of the criminals. Wang and Brown [51] used a spatio-
In this section, a review of the most relevant contributions to the temporal analysis for modeling criminal incidents, making use of a variety
literature is presented. of data types, such as spatial, temporal, geographic, and demographic
data. In a subsequent paper, Wang et al. [52] extended this prediction
2.1. Predictive policing model to include information proceeding from social network posts. A
similar approach is proposed by Gerber [23], nding that by combining
The term predictive policing is relatively recent and refers to the historical crime records with Twitter data from users in a specic geo-
application of quantitative techniques to foretell where crimes will take graphic area, the prediction performance improves for 19 of 25 crime
place in the short-term future. The National Institute of Justice (NIJ) de- types. Finally, Chen et al. [15] applied spatio-temporal analysis methods
ned it as taking data from disparate sources, analyzing them, and then to investigate patterns of offenses against property.
using the results to anticipate, prevent, and respond more effectively to
future crimes [36]. This technique is based upon advances in criminology, 2.2. The Police Districting Problem (PDP)
such as Hot Spot theories [43,54,53], and studies of the ecology of crime
[7,8]. Statistics based methods have been used since the release of District design can be seen as the problem of grouping the elementa-
CompStat in 1994, but it was only a few years ago that complex mathe- ry units of a given territory into larger districts, according to their
matical algorithms have been developed to address this problem in the relevant attributes. Depending on the problem faced, the attributes
M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 27

considered might belong to different contexts, including economic, de- hotspot approaches that apply GIS to combine crime rates and crash
mographic, geographic, and political contexts. In the last decades, the rates for prediction, aiming to reduce dispatch time. The main charac-
districting problem has been approached in a broad number of elds: teristics of the articles reviewed here are summarized in Table 1.
e.g., electric power, schools, electoral areas, and police patrols. A unied The research presented in this article innovates in the eld of DSS for
territorial design model that allows the formulation and solution of public security in several ways. First, most of the methodologies devel-
districting problems in a variety of applications is the subject of Kalcsics oped so far assume that the crime incidents can be represented as points
and Schreder [28], which also reviews the existing literature on territo- in time and space. As we elaborate in Section 3.1, this is not always the
rial design, highlighting application elds, criteria, and solution meth- case, as most of the time the victims are not aware of the time or the
odologies for these types of problems. location where the crime took place. Therefore we developed a method-
The Police Districting Problem (PDP) concerns the optimal ology that allows us to represent crime records having an indeterminate
partitioning of a territory into patrol sectors. Optimizing the congura- time or location. Second, the proposed methodology makes use of
tion of patrol sectors in the face of the expected crime activity in a classical time series models. Although not novel, these models have
particular shift or scenario provides a solid guideline for the effective two advantages: they are very good at capturing the seasonal compo-
deployment of the agents. Also, by concentrating the presence of police nents of the crime data and they do not require any additional spatial
agents in the areas with the highest expected levels of criminal activity, information (e.g., population density, average income, distance to
it fosters a shift of paradigm from detention to prevention, thanks to the risky locations), which makes them extremely applicable. Finally, to
deterrent effect of a police presence. The rst paper on PDP was by the best of the authors' knowledge, the DSS presented in this paper is
Mitchell [34], which proposes a clustering heuristic for the redesign of the rst to combine Predictive Policing capabilities with an optimization
patrol beats in Anaheim, California. The author considers the total ex- model that explicitly provides efcient partitions of the territory into
pected weighted distance to incidents, as well as a workload measure patrol sectors, rather than just a representation of the criminal hot spots.
dened as the sum of the expected service time and the expected travel
time. Bodily [6] adopts a utility theory model that incorporates the pref- 3. Structure of the P3-DSS
erences of three interest groups, namely, the citizens, the administra-
tors, and the service personnel. A simple local search algorithm swaps The DSS that we propose is composed of three main elements that
patrol beats from one sector to another to improve the value of the util- identify the predictive police patrolling strategies: Data Pre-Processing
ity function. Benveniste [1] was the rst author to include workload Unit (DPPU), Crime Risk Forecasting Unit (CRFU), and Patrol Sector
equalization in the optimization process, solving a non-linear stochastic Optimization Unit (PSOU).
model by means of an approximation algorithm. The model by D'Amico Fig. 1 illustrates the main loop of the P3-DSS, which shows how the el-
et al. [20] calculates sector workloads by calling an external software ements interact. Being a real-time system, the P3-DSS undertakes an in-
program, PCAM [12,13]. This external routine makes use of a queuing nite loop composed of two main parts. When a certain updating condition
model to compute the statistics regarding a sector, including the opti- is met (e.g., when a signicant number of records have been added to the
mal number of cars to be allocated. The simulated annealing algorithm crime reports database and a certain amount of time has passed since the
devised by D'Amico et al. [20] iteratively calls PCAM as a subroutine. last update), the system rst calls the DPPU to update the internal data
To enforce equity in terms of area, the ratio of the size of the largest structures and the CRFU to update the forecasting models. These opera-
and smallest sectors is bounded from above. Unlike the previous papers, tions are carried out in the background, and are invisible to the user.
Curtin et al. [19] apply a covering model to determine police patrol sec- Whenever a request for a patrolling conguration is sent by the user,
tors, so as to maximize the number of incidents that are close to the cen- the P3-DSS calls the PSOU and returns the resulting conguration. These
ters of the sectors. In a subsequent article [18], the authors include units will be presented in detail in the following subsections.
backup coverage (e.g., multiple coverage of high priority locations).
Zhang and Brown [57] propose a heuristic algorithm for the generation 3.1. Data Pre-Processing Unit (DPPU)
of districting, evaluated using an agent-based simulation model.
Contributions to the PDP have been extremely heterogeneous in Most of the research in predictive policing assumes that the criminal
terms of the objectives and methodologies adopted. As an example, a uni- incidents are associated with a determined point in time and space.
versal denition of workload has not been reached yet. A thorough review Ratcliffe [41] presents a methodology for the temporal description of
of all the research on the PDP, analyzing and comparing the different as- crime events where the time of incidence is indeterminate. However,
pects of the models and approaches proposed, can be found in [9]. many common crimes also have indeterminate spatial incidence,
e.g., pickpocketing. We now present a novel methodology for the
2.3. DSS for efcient policing spatio-temporal description of crime events that can be indeterminate
in both the temporal and the spatial dimensions.
The application of DSS to police environment has a growing role in The P3-DSS makes use of a three-dimensional data structure provid-
the literature. Some innovative strategies have been established to be ing a discretized representation of the space (i.e., the territory under the
effective and efcient when deciding on the best ofcer deployment jurisdiction of a district) and of time (i.e., the period of time considered
and schedule. In the late 1980s, Taylor and Huxley [47] proposed a in the historical data). In fact, we represent the territory under the juris-
scheduling system driven by shift turnovers, based on years of calls diction of a district as a grid, G, having I rows and J columns. The size of
data. This led to a declining response time compared to manual the grid cells can be determined by taking advantage of the results of
methods. However, in that study, different crime records were not con- Gorr and Harries [24], which show that the average monthly crime
sidered. More recently, Xue and Brown [56] proposed a DSS in which counts need to be on the order of 30 or more to achieve good forecast
criminal events were modeled as point processes. They intended to pre- accuracy. Time is discretized by considering the agents' shift as the
dict the spatial behavior of active criminals. By analyzing the offenders' time unit.1 The total number of time steps, T, can be easily computed
decisions, law enforcement would be empowered with better planning by calculating the number of shifts included in the period of time
and knowledge of the spatial patterns of crime. A few years later, anoth- encompassed by the historical data available.
er intuitive method discussed by Li et al. [33] suggested a model based The main data structure used by the DSS is a three-dimensional
on a fuzzy self-organizing map, identifying the characteristic of several array C, having dimension I J T. The value of each element,
crime patterns, to determine a better duty deployment. However, re-
cent research by Kuo et al. [32] criticized most programs for making 1
In the SNPC, shifts are scheduled as follows. Morning shift, from 8 AM to 3 PM; after-
use of a naive beforeafter evaluation method. The authors developed noon shift, from 3 PM to 10 PM; night shift, from 10 PM to 8 AM of the next day.
28 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

Table 1
References and structural characteristics of DSS for efcient policing.

Reference Data Objective(s) Technology Validated

Taylor and Huxley [47] Provided by: San Francisco Police Department, CA. Reduction of the cost of operations San Francisco Police Yes
Historical calls for service, time spent by call type, and increase of citizen safety and Department Computer Aided
percentage of calls requiring two or more police ofcer morale. (CAD) System
ofcers, and percentage of cars with two or more
ofcers allocated.
Xue and Brown [56] Provided by: Richmond Police Department, VA. Criminal Crime reduction Regional Crime Analysis No
incidents between July 1, 1997 and October 31, 1997. Program (ReCAP)
Includes more than 1200 crime observations.
Li et al. [33] Provided by: National Police Agency of Taiwan. Data Crime reduction Unknown No
originated from 20 county police bureaus in Taiwan from
2003 to 2004. Fourteen criminal categories were collected.
Kuo et al. [32] Provided by: College Station Police Department, TX. Crime Reduction of dispatching time and ArcGIS, KDE, Google Maps No
and crash data, from January 2005 to September 2010. number of crime and trafc events
Includes 65,461 offense reports and 14,712 crash reports.

ci,j,t , represents the number of crime reports associated with location Once C is built, we can use it to forecast the risk of crime in a specic
(i, j) G at time step t {1, , T}. The procedure executed by the DPPU shift.
to compute this value will be explained in the following.
The array C is initialized to 0. Next, each crime report is proportion- 3.2. Crime Risk Forecasting Unit (CRFU)
ally accounted for in the elements of the array involved in the criminal
event according to the following data included in the crime report The array C can be looked at from two different perspectives. In fact,
database of the SNPC: by selecting a specic time step, t f1; ; T g, the bi-dimensional matrix
C ;;t represents the distribution of crimes reported in the territory for
 
Event time window: Range of dates and times in which an event the selected shift. Similarly, by selecting a specic location i; j G,
occurred. the vector C i; j; is the time series of the crime counts for the selected
Event location: Place where the crime was committed, i.e., its
location. Since this number is an approximation to the real number of
geographical location. This might be specied as an address or, more
crimes committed, we can apply classical time series forecasting models
generally, as an area.
to predict the risk of crime for each cell of the grid. For instance, expo-
By proportionally, we mean that a crime is partially accounted for in
nential smoothing models assign progressively smaller weights (impor-
all elements ci,j,t referenced by the data. The following cases might
tance) to older data, whereas newer data is given progressively greater
occur:
weight. The use of classical time series forecasting models has also been
The time and location of the crime are known with certainty and are
validated by previous research on the topic. In fact, Gorr [25] and Cohen
limited to a single grid cell and time step (e.g., a robbery). The location
[17] agree that exponential smoothing models are very accurate at fore-
(i, j) and time step t can be determined unambiguously. The value of
casting crime series at the sub-district level. Following these results, the
ci,j,t is increased by 1.
CRFU considers for each location (i, j) G an exponential smoothing
The location of the crime is limited to a single grid cell and the time is
state space model [27]. As shown in Section 5, although this methodol-
expressed as a period of time covering more than one time step (e.g., a
ogy relies exclusively on crime location, the CRFU is capable of discern-
motor vehicle theft). In this case, the location (i, j) can be determined
ing the underlying pattern and producing good quality predictions.
unambiguously but the time is expressed as a range t, , t + n. Thus,
the value of the ci,j,t, , ci,j,t + n is increased according to the propor-
3.3. Patrol Sector Optimization Unit (PSOU)
tional part of the crime time range that falls into each time step.
The time of occurrence of the crime is contained in one time step but
Forecasting the risk of crime in an area is just the rst step toward
the location is not limited to a single grid cell (e.g., a breach of the
the denition of sound patrolling sectors in a district. By optimizing
peace). The time step t can be determined unequivocally but the loca-
the conguration of patrol areas, it is possible to focus resources on
tion is expressed as a set of locations {(i1, j1), , (in, jn)}. Therefore, the
the most relevant locations, with a consequential improvement in the
ci1 ; j1 ;t ; ; cin ; jn ;t are increased according to the proportional part of the
effectiveness of patrolling operations. After interviewing several service
area considered by the reports that falls into each grid cell.
coordinators and a number of agents involved in public safety opera-
The location is not limited to a single grid cell and the time of the
tions, several desirable characteristics were identied in order to nd
crime is contained in more than one time step (e.g., a pickpocketing
a good territory partition.
or evading a police car). In this case, rst all the elements involved
are identied, and then the value of each element is increased propor- Compact areas: a compact area allows for a better control of the terri-
tionally, as illustrated in the previous items. tory by the agents, as travel times from one point to another within

Fig. 1. The P3-DSS main loop.


M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 29

the area are minimal. Therefore, the more compact an area is, the sector that minimizes the sum of the distances to all the locations in
faster the response of agents who are in the area to emergency calls. the sector) is less than or equal to a dened constant, K. We recom-
Homogeneity in terms of workload: generating patrol sectors that are mend dening K as
similar in terms of workload is quite useful for two main reasons. First,


it ensures a more efcient distribution of work and, therefore, a better maxfI; J g
K p : 2
service to the public. Second, greater equality in workload increases p
the satisfaction of the agents.
Mutual support: It is desirable that agents be able to count on the The support received by a sector can be calculated by
support of agents assigned to other patrol sectors in case of need. n    o
  0 
b  s0 P dist os ; os K; s s0 ;
s
3
The mathematical optimization model proposed for the solution of
this problem partitions the area under the jurisdiction of a police district where os identies the median location of sector s and dist is the dis-
into a dened number of patrolling sectors, in the most efcient way. tance between two locations.2 The isolation of sector s is computed as
The resulting districting problem, called the Multi-Criteria Police s
Districting Problem (MC-PDP), has been presented in [9] and a fast p1b
s : 4
heuristic algorithm was proposed for its solution. Computational exper- p1
iments showed that the MC-PDP rapidly generates patrolling congura-
tions that are more efcient than those currently adopted by the SNPC. Risk, s. This attribute is a measure of the total risk associated to the
The main characteristics of the MC-PDP will be introduced in the sector that an agent patrols. It is expressed as the ratio of the total
following. risk of sector s, to the whole district risk.
X
ri j :
3.3.1. Input data and parameters i; js
X
s
: 5
The PSOU requires the following input data: r
i; jG i j

Let R be the bi-dimensional crime risk matrix for a future time step
t N T, having dimension I J. This matrix is computed by the CRFU Diameter, s. The diameter of a subset is dened as the maximum
by forecasting the crime risk level at each location (i, j) G. Thus, distance between any pair of locations belonging to that subset.
the value of every element ri,j R represents the predicted risk of It has been introduced in the MC-PDP as an efciency measure. In
crime at location (i, j) and time step t. fact, the diameter can be interpreted as the maximum distance
Let A be the bi-dimensional distance matrix, having dimension I J. that the agent associated to the district would have to travel in
The elements ai,j A are non-negative real numbers that represent case of an emergency call. Therefore, a small diameter results in a
the total length of the streets to be patrolled at location (i, j) G. low response time. The diameter measure used to evaluate a patrol
This matrix can be computed using the information provided in a sector is the ratio of the subset diameter to the maximum diameter
GIS. possible.
Let p be the number of patrolling sectors to be dened. We
assume p N 1. maxa;bs fdist a; bg
s : 6
Let w 4 be the vector of weights expressing the decision maker's maxa;bs fdist a; bg
preference associated with each attribute (see Section 3.3.3).
Let , 0 1 be the coefcient expressing the decision
maker's preference between optimization and workload balance By combining the attributes with the preference weights w dened
(see Section 3.3.4). by the decision maker, we can compute a measure of the workload
W s of a sector s as

W s w  s w  s w  s w  s : 7
3.3.2. Structure of a patrolling conguration
A feasible patrolling conguration is a partition P of the territory
considered. Each subset s P represents a patrol sector and is expressed
as a subset of locations, i.e., s G. From this point onward, the terms 3.3.4. Objective function
patrol sector and partition subset will refer to the same concept. According to the guidelines provided by the professionals of the
The number of subsets in the partition must be exactly p. All partition SNPC, the patrolling congurations should be as efcient as possible
subsets must be connected and convex. and, at the same time, they should distribute the workload homoge-
neously among the patrol sectors. Unfortunately, there might be a
3.3.3. Patrol sector attributes and workload trade-off between these requirements. The objective function of the
The MC-PDP evaluates the patrol sectors s P dened by a congu- MC-PDP takes into consideration the preferences of the decision
ration P according to four main attributes: area, isolation, demand, and maker for these factors.
diameter. All the attributes, explained in the following, are expressed X
  Ws
as dimensionless ratios, so as to be comparable. min ob jP  max W s 1  sP
: 8
sP p
s
Area, . This attribute is a measure of the size of the territory that
an agent should patrol. It is expressed as the ratio of the area The term maxs P{Ws} represents the worst workload, while the
encompassed by sector s, to the whole district area. Ws
term sP is the average workload. This objective function allows
X p
a the decision maker to examine the trade-off between optimization
i; js i j
X
s
: 1 and balance by a parametric analysis. In fact, by varying , the model
a
i; jG i j gives a range from optimization ( = 0) to balance ( = 1).

Isolation, s. In the MC-PDP, two districts support each other if the 2


Given the underlying grid structure, the distance function used in the MC-PDP is the
distance between their geometric medians (i.e., the location in a Manhattan distance.
30 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

Fig. 2. Operational envelope. Workload of the system as a function of the number of patrolling sectors p.

3.3.5. Solving the MC-PDP regarding the range of the impact of different partitioning strategies
The MC-PDP is an extremely complex model that cannot easily be using the same amount of resources, and the extent to which the
solved to optimality. In fact, not only does modeling the property of workload may be unnecessarily increased if suboptimal plans are imple-
subset connectivity make the MC-PDP intractable in large problems, mented. Prior examples of depicting similar envelopes in other applica-
but also no linear formulation for the convexity condition has been tion settings can be found in [16,31,50].
presented in the literature. Given that computational time is
criticalthe user expects a solution within a reasonable timethe
4.1. Computing the operational envelope
MC-PDP is solved by a Greedy Randomized Adaptive Search Proce-
dure (GRASP) algorithm. This methodology, thoroughly described
An algorithm for the computation of the operational envelope is
and tested in [9], is capable of generating in just one minute patrol-
presented in Algorithm 4.1.
ling congurations that are more efcient than those currently
adopted by the SNPC. Algorithm 1. Algorithm for the computation of the operational
envelope.

4. The operational envelope

The MC-PDP nds a good partition of the territory, according


to different attributes. When considering suboptimal ways of
subdividing a district, the basic question is what is the loss of ef-
ciency involved in terms of workload. We can approximate this
loss of efciency by calculating the distance from the best solution
value found. For a xed number of patrolling sectors and weights,
we can represent the increase in workload (or loss of system ef-
ciency) as shown in Fig. 2.
The operational envelope presented refers to Saturday, 10/13/2012,
night shift in the Central District of Madrid (see Section 5) and was
computed using the real crime distribution in the district. However,
we can obtain an approximate operational envelope by using the fore-
cast crime distribution provided by the CRFU. In Fig. 2, the values on
the x-axis represent the number of patrolling sectors. The values on
the y-axis display the workload, computed as in Eq. (8). For this illustra-
tive example, we assign to the weights and the balance coefcient the
following values, (w, w, w, w) = (0.45, 0.05, 0.45, 0.05) and =
0.1. As the number of patrolling sectors increases, the workload is
consequently decreased. Fig. 2 has two trends: the upper trend displays
the worst case workload (i.e., the workload obtained when the least ef- Computing (or approximating) the operational envelope for a turn is
fective patrolling plan is implemented) whereas the lower trend depicts computationally expensive. In fact, the lower trend points are the best
the best case workload (i.e., the workload when implementing the best solutions found by the MC-PDP, while the upper trend points are the
patrolling plan for different values of p). It can be easily seen that the greatest solutions found by a maximization version of the MC-PDP.
two trends dene a range of losses, from the best case to the worst For the determination of the operational envelope in Fig. 2, we ran the
case, that encompasses all the possible ways of protecting the district. optimization algorithms for each value of p iteratively and stopped the
This region is referred to as the operational envelope. Knowing the execution when there was no improvement in the best solution for 10
structure of the envelope can be helpful for patrol planning decisions. consecutive iterations. This stopping criterion is often applied in
The lower curve represents a situation of complete control. Thus, the exploration-based optimization algorithms such as Genetic Algorithms
optimal patrolling strategy can be devised. On the other hand, the to ensure that the method has converged [29,42]. In terms of computa-
upper curve shows the effects of applying the worst possible patrolling tional time, it took approximately two hours to compute the operational
scheme. The thickness of the envelope provides valuable information envelope presented in Fig. 2, which is quite time consuming. However,
M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 31

since the operational envelope is a strategical tool that needs to be


calculated only once per shift, this computational time is reasonable.

4.2. Calculating the efciency loss

The decision-maker can exploit the information provided by the op-


erational envelope to compare alternative patrolling congurations and
evaluate the potential benets of corrective plans, such as changes in
the number of patrolling sectors, or investments in the analysis of the
system to acquire a better denition of the data and the parameters.
In fact, the operational envelope can be used to compute the percentage Fig. 3. SNPC-GIS diagram. A line connecting two elements indicates bi-directional commu-
efciency loss associated to a certain partition P: nication between them.

objP obj P
Efficiency loss 100  9
objP obj P crime records.
GIS: The GIS in use in the SNPC allows visualizing crime records on the
where P+ and is the best and P the worst known partition for the set of map and representing the location of police vehicles in the territory.
attributes.

5. Case study: the Central District of Madrid The P3-DSS proposed in this paper is composed of three fundamental
parts that interact naturally with the existing SNPC-IS, as illustrated in
To test the P3-DSS, we developed an initial version considering the Fig. 4.
thefts reported during the years 2008 to 2012 in the Central District of The DPPU interacts with the ETL Module to access the Reports DB,
Madrid. The dataset includes exactly 105,755 incidents. This case get the crime records data, and to combine this data with the geograph-
study focuses on theft as it is the single most frequent type of crime ical information returned by the Map Server to build the crime risk
committed in Spain and one of the main priorities for the SNPC is its matrix C. This data is passed to the CRFU to forecast crime risk levels
reduction. However, extending the P3-DSS to other districts and to con- for future shifts. Both the DPPU and the CRFU need to regularly update
sider several types of crime is straightforward and can be accomplished their data structure and models so as to always have the best forecasting
with little change in the structure of the units and the models. Neverthe- quality. The PSOU obtains the forecast crime risk levels from the CRFU,
less, the nal version of the P3-DSS will require dedicated hardware to and the user's preferences from the GIS. Finally, the GIS is connected
be able to cover the whole national territory. to the CRFU and the PSOU to make queries regarding the distribution
of crime risk in a future shift and the recommended patrol sector cong-
5.1. Overview of the Central District of Madrid uration, respectively. The GIS has been updated to visualize the patrol
congurations, while the forecast crime risk levels are represented
Madrid is the capital of Spain and the most populous city in the using an existing feature for the display of heat maps.
country. The total population of the city in 2013 was 3,215,633 people,
and 6,369,162 people in the metropolitan area. The Central District of
Madrid, on which we focus our research, has an area of more than 5.3. Crime risk prediction quality
2 mile2, and is limited by the trafc circles of Segovia, Toledo, Valencia,
and Atocha. It is composed of six neighborhoods: Palacio, Embajadores, We now analyze the quality of the crime risk forecast given by the
Cortes, Justicia, Universidad, and Sol. Its population is approximately CRFU. The seasonality period-length was chosen according to prelimi-
150,000 people. The population of the Central District of Madrid is ex- nary experiments on the dataset that showed that the best performance
tremely heterogeneous; there is also a large transient population that is obtained when using a seasonality period-length of 21 shifts.
increasingly commutes to this district for reasons of work, sightseeing,
or leisure.
In Spain, the security of towns is the responsibility of the SNPC, usu-
ally sharing the territory with other local security forces. Currently, the
inspector in charge of civil protection operations in a shift decides the
distribution of agents in the district. This decision is normally taken
considering mostly their personal experience, their intuition, and also
some descriptive statistics, such as the summary of the criminal activity
of the last days.

5.2. Implementation and integration of P3-DSS

The DPPU and the CRFU have been developed in R [40] and then em-
bedded in C++. For their implementation, the following R packages
have been used: sp [37,4], rgeos [3], maptools [2], and forecast [26].
The PSOU has been programmed entirely in C++.
The GIS currently in use by the SNPC (SNPC-GIS) is structured as
shown in Fig. 3.
The SNPC-GIS is composed of the following elements:

Map Database and Server: These units provide access to the updated
road maps of Spain. Fig. 4. P3-DSS integrated into the SNPC Information System. A line connecting two
Reports Database and ETL Module: These units provide access to elements indicates bi-directional communication between them.
32 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

5.3.1. Dataset

0.24323
0.64897
5.97210
2.97830
0.09919
0.11087

1.53111

4.73182

0.59623
12/2012

235.93%

158.88%

537.77%
We assessed the quality of the CRFU forecasts by computing the
forecasting Mean Square Error (MSE), considering all the months in
2012 and three areas of interest in the Central District of Madrid. For
each validation period, we trained the forecasting model using all the
0.24023
0.48787
6.04007
1.54950
0.10040
0.02750

2.04720

3.81717

0.33392
criminal records prior to the period considered (e.g., for the January

419.62%

246.35%

1214.25%
11/2012

2012 validation period, we trained the forecasting models using the his-
torical data from January 2008 to December 2011). Next, we computed
the validation MSE for the following areas of interest of the district:
0.23104
1.04725
6.12261
1.89223
0.10124
0.05627

2.16645

3.96208

0.64930
Tribunal. Located in the north of the district, Tribunal is a well-
206.87%

209.39%

1153.90%
10/2012

frequented crossroad positioned right at the border between two


areas that are very popular for nightlife: Malasaa and Chueca.
Puerta del Sol. One of Madrid's main attractions and a prominent
meeting place for citizens and tourists alike, Puerta del Sol is a square
0.23081
0.25698
6.21466
1.16589
0.10210
0.04322

0.64585

1.71565

0.45586

located in the center of the district.


251.32%

147.15%

1054.74%
09/2012

El Rastro. An area located in the south of the district that hosts a


famous ea market every Sunday morning.
1.640898
0.23097
0.23633
6.30892
1.56276
0.10296
0.02678

0.40188

0.23759
08/2012

170.05%

100.05%

887.19%

5.3.2. Performance analysis


The training and validation MSEs are given in Table 2a. Except for
January 2012/Puerta del Sol, the CRFU performed extremely well in
0.134499

validation, with a maximum validation MSE of 2.98, corresponding to


0.23144
0.26313
6.39895
1.56948
0.10267

0.40318

3.00656

0.56963
07/2012

153.22%

191.56%

423.52%

an average forecasting error of 1.73 thefts per shift. The January


2012/Puerta del Sol case can be easily explained by observing Fig. 5.
In fact, the plot has a peak of more than 80 thefts in the rst shift.
This observation, that statistically can be considered an outlier, is due
0.23238
0.21446
6.49690
0.84760
0.10293
0.08519

0.56323

1.71924

0.50386

to the fact that the rst shift of January 2012 corresponds to New Year's
06/2012

262.63%

202.84%

591.45%

Eve. It is traditional in Madrid to celebrate this event in the Puerta del Sol.
Therefore, the high aggregation of people in the square and the festive
atmosphere contributed to the extraordinarily high number of thefts.
(b) Forecasting MSEs obtained using the baseline. The percentages are the ratios of the baseline MSEs to the CRFU MSEs.
0.23365
0.18121
6.60658
1.52356
0.10307
0.08281

0.51028

2.79284

0.41429
05/2012

281.60%

183.31%

500.29%

5.3.3. Comparison with the baseline


To understand the quality of the predictions returned by the CRFU
we compare the results obtained against a baseline. Table 2b presents
0.23004
0.41565
6.68529
2.69727
0.10295
0.11660

1.16131

4.47852

0.94083

the MSEs obtained by a baseline model that predicts tomorrow's


04/2012

279.40%

166.04%

806.89%

crime rate to be the same as today's crime rate. In absolute terms, the
baseline model performs fairly well, with relatively low values of valida-
tion MSE. We can conclude that, with the exception of the January
2012/Puerta del Sol case, the crime risk level is quite simple to forecast
0.23043
0.21372
6.78993
1.25833
0.10347
0.07405

1.23134

2.74614

0.50907
03/2012

using only information on time and location, in the dataset considered.


576.15%

218.24%

687.47%

In the table, the percentages are the ratios of the MSEs obtained by the
baseline model to those of the CRFU. These values conrm that the CRFU
provides a much better prediction than the baseline model. In fact, the
0.22439
0.53738
6.91950
1.27499
0.11009
0.26832

2.07518

2.68679

1.18519

validation MSE of the baseline model is always larger than the MSE of
02/2012

386.17%

210.73%

441.71%

the CRFU, especially in El Rastro, where it is always more than four


Forecasting MSEs obtained by the CRFU and by the baseline.

times larger.
134.45202
69.19553
0.22168
0.37128

0.10257
0.03804

1.48378

0.18613
5.5535
01/2012

399.64%

194.31%

489.30%
(a) Forecasting MSEs obtained by the CRFU.

PUERTA DEL SOL, JANUARY 2012


60
Validation

Validation

Validation

Validation

Validation

Validation

Thefts
Training

Training

Training

Ratio

Ratio

Ratio
MSE

0 20
*Puerta del Sol

*Puerta del Sol

0 20 40 60 80
*El Rastro

*El Rastro
*Tribunal

*Tribunal

Shift
Table 2

Area

Fig. 5. Crime distribution for January 2012.


M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 33

(a) CONF2: the city is divided into


two big sectors by the Gran Via, the
main street in the district, and the (b) CONF6: the district is par-
agents are free to patrol the assigned titioned according to its neighbor-
area ad lib. hoods.

Fig. 6. Patrolling congurations currently adopted by the SNPC in the Central District of Madrid. Each sector is represented by a different color. (For interpretation of the references to color
in this gure legend, the reader is referred to the web version of this article.)

5.4. Predictive police patrol congurations quality (viz., Universidad and Justicia) while the southern sector includes four
neighborhoods (viz., Palacio, Sol, Embajadores, and Cortes).
Camacho-Collados et al. [9] extensively analyzed the performance of
the MC-PDP in the presence of perfect information. We now assess the To be able to compare the performance of these congurations with
quality of the patrolling congurations generated by the P3-DSS based those determined by the optimization algorithm, we represented
on the forecast crime risk. CONF6 and CONF2 using the same grid structure adopted by the optimi-
zation algorithm, as illustrated in Fig. 6. The cells of the grid shared by
more than one sector have been assigned to the sector occupying
5.4.1. Comparison with standard congurations most of its area. It should be noticed that both congurations have one
We now analyze the quality of the solutions found by the optimiza-
sector that is not convex, i.e., the green sector in CONF6 and the light
tion algorithm by comparing them to the patrolling congurations blue sector in CONF2. Therefore, the congurations currently adopted
currently adopted by the SNPC.
by the SNPC would be infeasible according to the optimization model
In the Central District of Madrid, on an average day, one of the
proposed. This might result in better attribute values for these solutions
following patrol sector congurations is applied:
than those achievable with a feasible solution by the MC-PDP.
To assess the quality of the predictive patrol sector congurations,
we ran an analysis composed of the following steps:
CONF6: The district is partitioned according to its six neighborhoods.
CONF2: The district is split into two big sectors by the Gran Via, the 1. Generation of 1000 random test instances. Each instance is repre-
main artery in the territory, and the agents are free to patrol the sented as a tuple (w, w, w, w, , shift). For each sample, the attri-
assigned areas ad lib. The northern sector includes two neighborhoods bute weights (w, w, w, w) are sampled from a uniform

(a) SATT3: Saturday, 10/13/2011, (b) SUNT1: Sunday, 10/14/2011, (c) MONT2: Monday, 10/15/2011,
night shift (10 PM-8 AM). morning shift (8 AM-3 PM). afternoon shift (3 PM-10 PM).

Fig. 7. Maps of the number of thefts reported in the Central District of Madrid. The red shade represents a high crime level while the white shade represents no criminal activity. (For
interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)
34 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

(a) Prediction for SATT3: Satur- (b) Prediction for SUNT1: Sunday, (c) Prediction for MONT2: Monday,
day, 10/13/2011, night shift (10 PM- 10/14/2011, morning shift (8 AM-3 10/15/2011, afternoon shift (3 PM-
8 AM). PM). 10 PM).

Fig. 8. Maps of the forecast crime risk in the Central District of Madrid. The red shade represents a high crime risk level while the white shade represents no criminal risk. (For interpretation
of the references to color in this gure legend, the reader is referred to the web version of this article.)

distribution U(0,1) and normalized, the balance coefcient is sam- the P3-DSS produces patrolling congurations that dominate those
pled from a uniform distribution U(0,1) and the shift is chosen ran- currently adopted by the SNPC.
domly from the set of all the shifts in 2012.
2. Forecasting of the crime risk distribution and extraction of the valida- 5.4.2. Comparing the loss of performance
tion crime distribution for each of the test instances. It is difcult to understand what the real improvement is in terms of
3. Generation of patrolling congurations based on the forecast crime efciency resulting from using the P3-DSS. That is because the objective
risk by running the MC-PDP once, using 2 and 6 patrol sectors function (Eq. (8)) includes both a balance term and a global perfor-
(p = {2, 6}). The optimization algorithm has been run for 60 s to mance term. Fortunately, we can use the operational envelope (see
simulate a real-time environment. Section 4) to compare the patrolling congurations in terms of efcien-
4. Evaluation of patrolling congurations generated by the P3-DSS cy loss (Eq. (9)).
using the validation data. Given the computational time required to compute the operational
5. Evaluation of the standard patrolling congurations currently envelope for each case, this analysis has been limited to three shifts
adopted by the SNPC (CONF2 and CONF6) using the validation data. and one conguration of the attributes. The shifts and the attributes
6. Statistical comparison of the objective function values obtained on have been identied with the help of a service coordinator in charge
the same test instances by the standard congurations and those of the patrolling operations of the Central District of Madrid. In the
generated by the P3-DSS. course of many meetings, we asked the service coordinator to identify
a small number of shifts that could be considered as typical scenarios
The results of the statistical analysis will be presented in the follow- and that presented very different crime activity patterns. The following
ing. First we tested the data for each group for normality using a shifts are considered:
ShapiroWilk Normality Test and found that they did not follow a nor-
mal distribution. Then, we applied a Friedman Test, a nonparametric SATT3: Saturday, 10/13/2012, night shift (10 PM8 AM).
test of nonindependent data from two or more groups that does not SUNT1: Sunday, 10/14/2012, morning shift (8 AM3 PM).
require the data to proceed from a normal distribution. The statistical MONT2: Monday, 10/15/2012, afternoon shift (3 PM10 PM).
difference is signicant in both cases. In fact, the p-values are p =
2.56e 12 for the 2 patrol sectors case, and p = 0.02781 for the 6 patrol
sectors case. We can therefore state that, even in the face of uncertainty, These three shifts were chosen for their representativeness. Fig. 7
illustrates the distribution of thefts in these three shifts. SATT3 is charac-
Table 3 terized by a high level of nightlife, with people coming from other
Comparison of the objective function values.
districts of Madrid as well as other cities. In the picture it can be seen
Shift Conguration p Obj (P) that thefts are committed in almost all the territory, with the highest
(a) Objective function values of the patrol sector congurations currently adopted levels concentrated around Plaza Callao, a busy meeting place, Plaza
by the SNPC. Mayor, the central plaza of the city, and Lavapies, a difcult area.
*SATT3 CONF2 2 0.55086 SUNT1 has a moderate level of criminality, mostly concentrated in the
CONF6 6 0.20655
south of the district where the popular El Rastro ea market is held
*SUNT1 CONF2 2 0.56515
CONF6 6 0.21236
*MONT2 CONF2 2 0.55074 Table 4
CONF6 6 0.20830 Comparison of the patrol sector congurations obtained by P3-DSS with those currently
adopted by the SNPC.
Dataset p Obj (P) 95% CI
Dataset Conguration p SNPC solutions P3-DSS solutions efciency
(b) Objective function values of the patrol sector congurations obtained by
efciency loss (%) loss (%) 95% CI
P3-DSS. 95% condence intervals over 50 runs.
*SATT3 2 [0.54560,0.54560] *SATT3 CONF2 2 15.77 [0,0]
6 [0.19699,0.20061] CONF6 6 20.01 [10.15,14.38]
*SUNT1 2 [0.54695,0.54695] *SUNT1 CONF2 2 82.75 [0,0]
6 [0.19703,0.19973] CONF6 6 24.32 [8.81,12.04]
*MONT2 2 [0.54112,0.54112] *MONT2 CONF2 2 24.14 [0,0]
6 [0.19399,0.19708] CONF6 6 21.46 [7.12,10.74]
M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 35

(a) (b)

Fig. 9. Best patrolling congurations identied by the P3-DSS for the shift SATT3 with 2 and 6 patrol sectors. Each sector is represented in a different color. (For interpretation of the
references to color in this gure legend, the reader is referred to the web version of this article.)

every Sunday morning. Finally, MONT2 presents the characteristics of a Table 3b. As expected, the condence intervals are better (lower) than
normal business day, with low levels of criminal activity, mostly the objective function values of the corresponding SNPC congurations
concentrated in the commercial area. in all cases.
Fig. 8 shows the forecast crime risk distribution obtained by the To understand the real improvement in terms of efciency resulting
CRFU. Visually, the real and forecast crime distributions appear to be from the adoption of the P3-DSS, we compute the efciency loss for both
very close. types of patrolling congurations.
Concerning the weights and the balance coefcient, we explained Table 4 provides the efciency loss, computed according to Eq. (9),
the meaning of each attribute in the model to the service coordinator. associated to the patrolling congurations currently adopted by the
Then we asked her to rate from 0 to 10 the importance of each attribute SNPC and the 95% condence interval of the efciency loss relative to
and her preference between patrolling efciency and workload balance. the solutions identied by the P3-DSS. For this experiment the opera-
The following values were obtained after normalizing and scaling the tional envelope was approximated by running the optimization algo-
gures provided: rithms on the validation data. The results in the table show that there
is a signicant improvement in terms of efciency when implementing
Attribute weights, (w, w, w, w) = (0.45, 0.05, 0.45, 0.05). the congurations identied by the P3-DSS. These results conrm the
Balance coefcient, = 0.1. usefulness of the P3-DSS as a policing DSS for the data and the parame-
ters considered in the experiments.
Figs. 911 show the best patrolling congurations obtained by the
The objective function values (Eq. (8)) of the standard patrolling
P3-DSS for the three shifts considered. We can see that these patrol sec-
congurations are provided in Table 3a.
tors have signicant differences from those currently in use (Fig. 6).
The MC-PDP was run for 60 s using the parameters provided by the
Some insights will be given in the following:
service coordinator. As the optimization algorithm is random in nature,
we ran each conguration 50 times, using the forecast data. After SATT3: Police activity is focused on the Gran Via, the main artery of the
that, we evaluated the solutions using the validation data. The 95% con- city that runs from the top-left corner of the district to the center, and
dence intervals of the solution values using the real data are shown in then goes toward the east. The reason for that is that the Gran Via and

(a) (b)

Fig. 10. Best patrolling congurations identied by the P3-DSS for the shift SUNT1 with 2 and 6 patrol sectors. Each sector is represented in a different color. (For interpretation of the
references to color in this gure legend, the reader is referred to the web version of this article.)
36 M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537

(a) (b)

Fig. 11. Best patrolling congurations identied by the P3-DSS for the shift MONT2 with 2 and 6 patrol sectors. Each sector is represented in a different color. (For interpretation of the
references to color in this gure legend, the reader is referred to the web version of this article.)

its surroundings are very popular nightlife areas. would be (1) the denition of a system for the objective evaluation of
SUNT1: The patrolling conguration concentrates on the southern the effectiveness of an agent and (2) a model capable of determining
part of the district, where most of the crimes happen on Sunday morn- the minimum number of agents required in a specic future shift.
ing because of the popular ea market. We hope that this work will be a useful source of ideas for future
MONT2: The city is uniformly partitioned between north-east and research and will foster and encourage more collaborations between
south-west. The conguration with 6 patrol sectors assigns higher police agencies and researchers in the elds of computer science, oper-
importance to the central-western part of the district, corresponding ations research, mathematics, and statistics.
to the commercial area.
Acknowledgments

The authors would like to thank the Spanish National Police Corps
6. Conclusions personnel of the SEYCO (Statistics and Operative Control Section), the
CPD (Data Processing Center), the Sala 091 of Madrid (Police Emergen-
This study represents a rst step toward a comprehensive system cy Call Center), the Central District of Madrid, and the Central District of
based on mathematical algorithms, providing an objective and effective Granada for their help and collaboration. The investigation of Camacho-
approach to decision-making for public security police ofcers. Speci- Collados was partially nanced by the Spanish Police Foundation re-
cally, we made use of predictive techniques for controlling the variabil- search grant for civil servants of the Spanish National Police Corps and
ity of crime in sufciently small spatial areas, to subsequently develop a by the Fulbright Program. The research of Liberatore was supported
tool to help in the management of a proper geographical distribution of by the Spanish Government, grant TIN2012-32482. All support is grate-
the available patrols in a concrete police shift. To evaluate this new ap- fully acknowledged.
proach, we applied our models to predict property crimes in the Central
District of Madrid. Based on our assessments with the real criminal inci- References
dent data, our models can not only predict future incidents accurately
[1] R. Benveniste, Solving the combined zoning and location problem for several emer-
but also determine a very efcient police distribution in a district.
gency units, Journal of the Operational Research Society 36 (5) (1985) 04330450.
Other contributions made by this paper are a methodology for the [2] R. Bivand, N. Lewin-Koh, maptools: Tools for Reading and Handling Spatial
temporal and spatial description of indeterminate crime events, and ObjectsURL: http://CRAN.R-project.org/package=maptools,2014 (R package ver-
an algorithm for the computation of the operational envelope of the sion 0.829).
[3] R. Bivand, C. Rundel, rgeos: Interface to Geometry Engine-Open Source (GEOS)URL:
problem under analysis. http://CRAN.R-project.org/package=rgeos,2014 (R package version 0.34).
Despite the very promising results obtained, there is still room for [4] R.S. Bivand, E. Pebesma, V. Gomez-Rubio, Applied Spatial Data Analysis with R,
improvement. In fact, the SNPC is currently enhancing the Reports DB Second edition Springer-Verlag, 2013. (URL: http://www.asdar-book.org/).
[5] R.L. Block, C.R. Block, Crime and Place, chapter Space, Place and Crime: Hot Spot
to support the geolocation of the events, rather than relying on written Areas and Hot Places of Liquor-Related Crime, Criminal Justice Press/Willow Tree
descriptions. Also, the forecasting models in the CRFU might be extended Press, 1995. 145183.
to include additional variables such as weather conditions (e.g., amount [6] S. Bodily, Police sector design incorporating preferences of interest groups for equal-
ity and efciency, Management Science 24 (12) (1978) 01301.
of rain, temperature), special events (e.g., demonstrations, sport events, [7] P.J. Brantingham, P. Brantingham, Patterns in Crime, Prentice-Hall, 1984.
parades), and social network posts (e.g., Twitter, Instagram). The addi- [8] R.J. Bursik, H.G. Grasmick, Neighborhoods and crime: the dimensions of effective
tion of these and other correlated variables should result in better community control, Lexington Books, 1993.
[9] M. Camacho-Collados, F. Liberatore, J.M. Angulo, A multi-criteria police districting
crime risk predictions. Finally, the optimization model in the PSOU problem for the efcient and effective design of patrol sector, European Journal of
could be adapted to consider census districts instead of the grid struc- Operations Research, (2015)http://dx.doi.org/10.1016/j.ejor.2015.05.023.
ture. Apart from eliminating the issue of determining the size of the [10] J.M. Caplan, L.W. Kennedy, Risk Terrain Modeling Compendium, Rutgers Center on
Public Security, 2011.
grid cells, this would simplify the exchange of information and data
[11] J.M. Caplan, L.W. Kennedy, J. Miller, Risk terrain modeling: brokering criminological
among public agencies. theory and GIS methods for crime forecasting, Justice Quarterly 28 (2) (2011)
This research represents a rst step toward the implementation of a 03600381.
comprehensive DSS for police corps that provides support for all the [12] J. Chaiken, P. Dormont, A patrol car allocation model: background, Management
Science 24 (12) (1978) 0128001290.
operations carried out by police agencies. According to some agents in [13] J. Chaiken, P. Dormont, A patrol car allocation model: capabilities and algorithms,
charge of public security in the SNPC, the most interesting features Management Science 24 (12) (1978) 0129101300.
M. Camacho-Collados, F. Liberatore / Decision Support Systems 75 (2015) 2537 37

[14] S. Chaturapruek, J. Breslau, D. Yazdi, T. Kolokolnikov, S.G. McCalla, Crime model- RAND Corporation (2013) http://www.rand.org/pubs/research_reports/RR233.
ing with Levy ights, SIAM Journal on Applied Mathematics 73 (4) (2013) Also available in print form.
0170301720. [39] A.J. Pinizzotto, E.F. Davis, C.E. Miller III., Emotional/Rational Decision Making in Law
[15] P. Chen, H. Yuan, D. Li, Spacetime analysis of burglary in Beijing, Security Journal 26 Enforcement, Technical report, Federal Bureau of Investigation, 2004.
(1) (2013) 115. [40] R. Core Team, R: a Language and Environment for Statistical Computing, R
[16] R.L. Church, M.P. Scaparra, Analysis of facility systems reliability when subject to at- Foundation for Statistical Computing, Vienna, Austria, 2014. (URL: http://
tack or a natural disaster, in: A.T. Murray, T.H. Grubesic (Eds.), Critical Infrastructure: www.R-project.org/).
Reliability and Vulnerability: Advances in Spatial Science, Springer-Verlag 2007, [41] J.H. Ratcliffe, Aoristic signatures and the spatio-temporal analysis of high volume
pp. 221241. crime patterns, Journal of Quantitative Criminology 18 (1) (2002) 343.
[17] J. Cohen, Development of Crime Forecasting and Mapping Systems for Use by Police, [42] M. Safe, J. Carbadillo, I. Ponzoni, N. Brignole, On Stopping Criteria for Genetic Algo-
Technical report, U.S. Department of Justice, 2006. rithms, in: A.L.C. Bazzan, S. Labidi (Eds.), Advances in Articial Intelligence SBIA
[18] K. Curtin, K. Hayslett-McCall, F. Qiu, Determining optimal police patrol areas with 2004, volume 3171 of Lecture Notes in Computer Science, Springer-Verlag 2004,
maximal covering and backup covering location models, Networks and Spatial pp. 405413.
Economics 10 (1) (2010) 125145. [43] L.W. Sherman, P.R. Gartin, M.E. Buerger, Hot spots of predatory crime: routine activ-
[19] K. Curtin, F. Qui, K. Hayslett-McCall, T. Bray, Geographic Information Systems and ities and the criminology of place, Criminology 27 (1) (1989) 2756.
Crime Analysis, Chapter Integrating GIS and Maximal Coverage Models to Determine [44] M.B. Short, A.L. Bertozzi, P.J. Brantingham, Nonlinear patterns in urban crime:
Optimal Police Patrol Areas, Idea Group Inc., 2005. 214235. hotspots, bifurcations, and suppression, SIAM Journal on Applied Dynamical
[20] S. D'Amico, S. Wang, R. Batta, C. Rump, A simulated annealing approach to police Systems 9 (2) (2010) 462483.
district design, Computers & Operations Research 29 (6) (2002) 667684. [45] M.B. Short, M.R. D'Orsogna, V.B. Pasour, G.E. Tita, P.J. Brantingham, A.L. Bertozzi, L.B.
[21] Z. Friend, Predictive policing: using technology to reduce crime, FBI Law Enforce- Chayes, A statistical model of criminal behavior, Mathematical Models and Methods
ment Bulletin (2013) URL: http://leb.fbi.gov/2013/april/predictive-policing-using- in Applied Sciences 18 (Suppl.:0) (2008) 12491267.
technology-to-reduce-crime. [46] M.A. Smith, D.E. Brown, Application of discrete choice analysis to attack point
[22] V. Furtado, A. Melo, A.L.V. Coelho, R. Menezes, R. Perrone, A bio-inspired crime patterns, Information Systems and e-Business Management 5 (3) (2007) 255274.
simulation model, Decision Support Systems 48 (1) (2009) 282292. [47] P.E. Taylor, S.J. Huxley, A break from tradition for the San Francisco Police: patrol of-
[23] M.S. Gerber, Predicting crime using Twitter and kernel density estimation, Decision cer scheduling using an optimization-based decision support system, Interfaces 19
Support Systems 61 (1) (2014) 115125. (1) (1989) 424.
[24] W. Gorr, R. Harries, Introduction to crime forecasting, International Journal of [48] G.E. Tita, J. Cohen, J. Engberg, An ecological study of the location of gang set space,
Forecasting 19 (4) (2003) 551555. Social Problems 52 (2) (2005) 272299.
[25] W. Gorr, A. Olligschlaeger, Y. Thompson, Short-term forecasting of crime, Interna- [49] G.E. Tita, G. Ridgeway, The impact of gang formation on local patterns of crime,
tional Journal of Forecasting 19 (4) (2003) 579594. Journal of Research in Crime and Delinquency 44 (2) (2007) 208237.
[26] R.J. Hyndman, forecast: Forecasting Functions for Time Series and Linear Models [50] D. Urban, T. Keitt, Landscape connectivity: a graph theoretic perspective, Ecology 82
URL: http://CRAN.R-project.org/package=forecast, 2014 (R package version 5.4). (5) (2001) 12051218.
[27] R.J. Hyndman, A.B. Koehler, J.K. Ord, R.D. Snyder, Forecasting with Exponential [51] X. Wang, D.E. Brown, The spatio-temporal modeling for criminal incidents, Security
Smoothing: the State Space Approach, Springer-Verlag, 2008. Informatics 1 (0 2) (2012).
[28] J.N. Kalcsics, M. Schreder, Towards a unied territorial design approach applica- [52] X. Wang, M.S. Gerber, D.E. Brown, Social Computing, Behavioral Cultural Model-
tions, algorithms and GIS integration, TOP 13 (1) (2005) 174. ing and Prediction, chapter Automatic Crime Prediction Using Events Extracted
[29] N. Karma, C.Y. Suen, P.F. Guo, Palmprints: a novel co-evolutionary algorithm for from Twitter Posts, Springer-Verlag, 2012. 231238.
clustering nger images, Genetic and Evolutionary Computation GECCO 2003, [53] D. Weisburd, L. Green, Policing drug hot spots: the Jersey City drug market analysis
volume 2723 of Lecture Notes in Computer Science, Springer-Verlag 2003, experiment, Justice Quarterly 12 (4) (1995) 711736.
pp. 322331. [54] D. Weisburd, L. Maher, L. Sherman, Contrasting Crime General and Crime Specic
[30] L.W. Kennedy, J.M. Caplan, E. Piza, Risk clusters, hotspots, and spatial intelligence: Theory: The Case of Hot Spots of Crime, volume 4 of Advances in Criminological
risk terrain modeling as an algorithm for police resource allocation strategies, Theory, Transaction Press, 1992.
Journal of Quantitative Criminology 27 (3) (2011) 339362. [55] D. Weisburd, S. Mastrofski, A.M. McNally, R. Greenspan, J. Willis, Reforming to pre-
[31] H. Kim, M. O'Kelly, Survivability of commercial backbones with peering: a case study serve: Compstat and strategic problem solving in American policing, Criminology
of Korean networks, in: A.T. Murray, T.H. Grubesic (Eds.), 51st Annual North and, Public Policy 2 (3) (2003) 421456.
American Meetings of the Regional Science Association International, Springer- [56] Y. Xue, D.E. Brown, Spatial analysis with preference specication of latent decision
Verlag 2004, pp. 107128. makers for criminal event prediction, Decision Support Systems 41 (3) (2006)
[32] P.F. Kuo, D. Lord, T.D. Walden, Using geographical information systems to organize 560573.
police patrol routes effectively by grouping hotspots of crash and crime data, Journal [57] Y. Zhang, D.E. Brown, Police patrol districting method and simulation evaluation
of Transport Geography 30 (1) (2013) 138148. using agent-based model & GIS, Security Informatics 2 (7) (2013).
[33] S. Li, S. Kuo, F. Tsai, An intelligent decision-support model using FSOM and rule ex- [58] J.R. Zipkin, M.B. Short, A.L. Bertozzi, Cops on the dots in a mathematical model of
traction for crime prevention, Expert Systems with Applications 37 (10) (2010) urban crime and police response, 19 (5) (2014) 14791506.
71087119.
[34] P.S. Mitchell, Optimal selection of police patrol beats, The Journal of Criminal Law,
Miguel Camacho Collados is a mathematician, a police inspector of the Spanish National
Criminology and Police Science 63 (4) (1972) 577.
Police Corps, and a PhD candidate of the University of Granada. Currently he is a Fulbright
[35] G. Mohler, M. Short, P. Brantingham, F. Schoenberg, G. Tita, Self-exciting point
Grant holder and he is visiting the Department of Mathematics of the UCLA as an exchange
process modeling of crime, Journal of the American Statistical Association 106
scholar under the supervision of Prof. Bertozzi.
(493) (2011) 100108.
[36] NIJ, Predictive Policing, http://www.nij.gov/topics/law-enforcement/strategies/
Federico Liberatore received his PhD in Operational Research in 2010 from the University
predictive-policing/Pag es/welcome.aspx,2014.
of Kent, UK. He is a Fulbright Grant holder. His research interests include location analysis,
[37] E.J. Pebesma, R.S. Bivand, Classes and methods for spatial data in R, R News 5 (2)
fortication/interdiction models in networks, humanitarian logistics, policing models, and
(2005) (URL: http://cran.r-project.org/doc/Rnews/).
articial intelligence in games.
[38] W.L. Perry, B. McInnis, C.C. Price, S. Smith, J.S. Hollywood, Predictive Policing: The
Role of Crime Forecasting in Law Enforcement Operations. Santa Monica, CA:

You might also like