Advertisement

Distributed Agent-Based Simulation and GIS: An Experiment with the Dynamics of Social Norms

  • Nicola Lettieri
  • Carmine SpagnuoloEmail author
  • Luca Vicidomini
Conference paper
Part of the Lecture Notes in Computer Science book series (LNCS, volume 9523)

Abstract

In the last decade, the investigation of the social complexity has witnessed the rise of Computational Social Science, a research paradigm that heavily relies upon data and computation to foster our understanding of social phenomena. In this field, a key role is played by the explanatory and predictive power of agent-based social simulations that are showing to take advantage of GIS, higher number of agents and real data. We focus GIS based distibuted ABMs. We observed that the density distribution of agents, over the field, strongly impact on the overall performances. In order to better understand this issue, we analyzes three different scenarios ranging from real positioning, where the citizens are positioned according to a real dataset to a random positioning where the agent are positioned uniformly at random on the field. Results confirm our hypothesis and show that an irregular distribution of the agents over the field increases the communication overhead. We provide also an analytic analysis which, in a 2-dimensional uniform field partitioning, is affected by several parameters (which depend on the model), but is also influenced by the density distribution of agents over the field. According to the presented results, we have that uniform space partitioning strategy does not scale on GIS based ABM characterized by an irregular distribution of agents.

Keywords

Distributed agent-based social simulation GIS D-Mason Parallel computing Distributed systems ABM GIS 

1 Introduction

In the last decade, the investigation of the social complexity has witnessed a deep change from both theoretical and methodological standpoint. The emerging research area of Computational Social Science (CSS) is promoting a new scientific paradigm that heavily relies upon the power of data and computation to foster the understanding of social phenomena [7, 27]. The computationally inspired approach to the study of human societies is not only offering deeper insights into social reality, but is also supporting the design of more effective and contextualized policies [34]. Indeed, many of the problems policy makers have to cope with (spanning from economic instability to the spread of epidemics) are characterised by the nonlinearity and unpredictability of complex systems and CSS tools and methodologies are showing the power of shedding new light into the core mechanics of these problems [23].

In this scenario, a key role is played by agent-based social simulation models [1, 19] that represent an innovative way to investigate how local interactions taking place at the micro level between individuals can generate emergent properties of social structures. ABM has so far allowed to study the emergence of many collective phenomena and behaviours from cooperation [3, 30], to reciprocity [24] and social norms [28] with promising results for social science and policy modeling.

More Realistic (Complex) Is Better? The design of a social simulation model is an intricate task [20]. Social dynamics are the result of complex structures of interactions that involve at different levels individual cognition and behaviour, groups, institutions and the surrounding environment. The modeling enterprise implies the operational description of factors involved in generating the macro phenomenon under investigation. Even if interesting insights into social phenomena have been yielded thanks to very simple models (see, among the others, the famous Schelling’s segregation model [33]), agent-based simulation is nowadays increasingly expected to reach higher levels of realism in order to increase the explanatory and predictive power of models. This expectation implies different consequences: building agents with more complex cognitive architectures mimicking real psychological processes; explicitly representing space; reach an adequately high number of agents when simulating scale-dependent phenomena; linking simulation model to real world data.

The effects of this choice unfold at the same time on a scientific and technical level. As recently highlighted [9, 10] reaching a higher level of realism in social requires not only a strongly interdisciplinary modeling approach, but also the use of tools allowing to tackle the computational weight of more complex models. According to [29] parallelization is strictly necessary to run massive simulations that are needed to model meaningful complex social and economic phenomena: HPC is not simply a solution to speed-up the execution of simulation experiments, but also a way to enable completely new research questions.

In the next sections we focus on some technical issues arising when dealing with simulations with a high number of agents, real data, and GIS based environment. Our goal is twofold: analyse computational and programming issues arising when adding complexity to a relatively simple social simulation model; show how distributed computing can support advances in the investigation of social issues. We’ll do it implementing a distributed simulation drawing inspiration from a well known model of norm innovation dynamic [2], a very relevant topic for many areas of social science from economics to law. We will extend the original model putting agents endowed with the cognitive architecture of social norm recognition into a simulation setting with millions individuals, explicit space and data coming from official census.

Agent-Based Model and Geographical Information Systems. The ABM community has provided several tools to build ABM simulation. Many of these tools allow the developer to use GIS data in ABM simulation.

The GIS [21] (Geographic Information Science or Geographic Information Systems) term refers to a set of theories and techniques (especially computer-based) that allows benefit geographical data and metadata in the modeling ABM. The application of GIS data in the field of ABM is relatively recent, but the interest in this field led to the creation of dedicated community [22] and as described in the chapter four of a recent book “Geocomputation: a practical primer” [6] the interest in this field is intended to grow.

Many ABM tools support the GIS data. Among them we have:

Mason [4] , a discrete-event simulation core and visualization library written in Java, designed to simulate a wide range of ABMs. MASON is designed and maintained by George Mason University’s Evolutionary Computation Laboratory and the GMU Center for Social Complexity. Mason provides support for GIS data in an additional library named GeoMason [35].

GeoMason follows the same Mason design philosophy of being lightweight, modular, and efficient. GeoMason represents the basic GIS data in a geometric shapes supported via the JTS (Java Topology Suite API), which allows geometries related operations. GeoMason supports the ESRI [18] shape files providing a GeomVectorField Java object that represents the GIS data in the memory, and provides functionalities to access to geometries in order to read geospatial metadata and obtain geospatial positions of the objects.

Netlogo [36], a multi-agent programmable modeling environment. It is developed at the The Center for Connected Learning (CCL) and Computer-Based Modeling at Northwestern University. Netlogo provides an extension to support GIS data field. The extension allows the programmer to load vector GIS data (points, lines, and polygons), and raster GIS data (grids) into their models. Netlogo extension also supports the ESRI shapefiles.

Parallel and Distributed Multi Agent-Based Simulation. As described in [32] the scientific community have produced several tools and framework to run multi-agents system in a distributed environment:

D-Mason [11, 12, 17] is a distributed version of Mason library for running ABMs on distributed systems. Currently, D-Mason allows to parallelize simulation based on geometric fields and does not support the parallelization of GeoMASON. D-Mason allows to choose different communication layers: a layer based on a centralized strategy, that use the message broker Apache ActiveMQ and one based on the MPI standard; both layers adopt the Publish/Subscribe paradigm.

Flame [26] is designed to support lots of ABM. It allows parallelization by using the MPI standard to ensure the communication between the nodes. Flame is developed by the University of Sheffield and support also the GIS data.

RepastHPC [8] is the parallel version of Repast. RepastHPC is developed by the Argone institute of USA. RepastHPC uses MPI for the communications among the HPC environments. The current version RepastHPC supports GIS data.

In this paper, we perform several GIS-based ABMs on D-Mason. We use the distributed field of D-Mason in order to distribute the computation among the nodes. We notice that, we exploit GeoMason as static field that does not change during the simulation.

2 Experiment

In this section, we describe our experiment of distributed ABM GIS model implemented in D-Mason. In the following we describe the GIS data used to support the modeling of the population behaviors in the Italy region Campania. Then we provide anof our experimental ABM GIS model and the implementation of the model in D-Mason.

Model Space Representation. The environment in ABM [5] is not only a particular property of the model, but could be a relevant entity to understand the complex behavior of natural and artificial systems. Interesting features of ABM, compared to others modeling tools, concerns the interactions between the agents that do not take place in a vacuum, but happen in a structured environment that could influence and could be influenced by the agents interactions. These structured environments are named fields and can be social and physical environments (fields based on mathematical modeling like matrix), but also more complex structure like networks.

The environment representations is really crucial in order to address real-world problems (e.g., simulating the segregation in a particular area or simulating emergency strategies in natural disaster [16]). In this work we used a set of GIS “Campania dataset” data to support a model to simulate a toy experiment, inspired by a cognitive architecture model [6].

GIS Campania Dataset. The open-data platform of the “Regione Campania” provides the GIS dataset [31] about its region. The dataset is a ShapeFile ESRI shape on the geographical coordinate system WGS84 UTM zone 32 N and provides the subdivision of the region in geographical zones identified by a unique identifier.

Model Agents Movement Representation. A single agent in our model is a citizen living in Campania that has to travel every day to his work/study place. Agents behavior is based upon public available data released by ISTAT [25] (the Italian National Institute of Statistics), produced after the national population census made in October 2011.

The ISTAT’s table contains information about 2.5 millions citizens living in Campania, which each day travel to work (or study) and then go back home. It contains the following data: city of residence, gender, reason to travel (work or study), city of work/study, vehicle (on foot, by car, by train, etc.), time of departure, travel duration.

Model Description. The model we are going to show is a simplified version of the cognitive model designed by Andrighetto et al. in [2]. In our model, agents move in a representation of the Campania environment: agents’ home and working places are placed on the Campania map according to real data from ISTAT.

The cognitive aspect of the model follows. There is a set of norms (graphically depicted using different colors). For each norm, agents will hold a salience (a value in the [0,1] interval) that represents how much that norm is important for the agent. The norm that is the most important for an agent, will characterize the agent’s color (belief of the norm).

Agents continuously interact with neighbor agents trying to spread their opinion (color). When an agent meet an agent advertising a particular color, it will increase the salience for that specific norm. Agents are influenced by neighbors throughout the day, but with different weight depending on agent’s state (traveling, staying at home, staying at work/study). Of course, saliences will naturally decay over time.

Campania is divided into five provinces: Naples, Salerno, Benevento, Avellino and Caserta. We imagined that in each province of Campania there is a norm that is prevalent (i.e. 80% of the inhabitants are of that color). For instance: Naples is mostly Red; Salerno is mostly Green; Avellino, Benevento, Caserta are mostly Blue. In the remaining 20% of the population, the 15% will be Yellow (the color not advertized by any region), while the last 5% will be of a random color that is different from the region’s color. Back to our example, 80% of people in the Salerno province will be Green; 15% will be Yellow, and the remaining 5% will be a random chosen between Red and Blue.

The model simulates an entire day (24 hours) starting from midnight. When the simulation start, agents are staying at their home. Time of departure, travel duration, time of stay at work/study all depend on ISTAT data. Times and durations need to be converted into simulation steps: for instance, if travel duration for the agent is from 16 to 30 min, the simulation will assign a random duration in that interval. This duration will be converted in a number of steps, according to discretization time of the simulation (see Sect. 3.1).

The size of the simulated field and discretization time have a significant impact on the performances (in terms of efficiency) of the simulation. As we will, in Sect. 2, an agent moves at a speed that is calculated dividing the travel distance by the travel duration (simulation steps). This gives the speed of an agents, that is the maximum distance covered in a single simulation step. The largest agents’ travel distance is called maximum agents ride (\(\alpha \)) and will require a certain number of steps (maximum number of steps to perform a ride, \(\beta \)). So we can compute the maximum speed (\(\alpha /\beta \)): this parameter has a strong impact on the distributed model performances (the smaller the better).

In order to evaluate the performances of the distributed simulation framework D-Mason we also included two further modifications to our model, concerning the way agents are placed on the map. We will refer to the model we just described as the one with Real positions; agents are place according the real population density. A second model, called CRandom, places agents uniformly at random on the entire Campania territory. The latest model, called Random, places agents uniformly random on a 2D continuous space. The latest model represents the best case for distribution, as it allows the model to balance the workload on multiple LPs (Logical Processors). It is, although, very unrealistic. The CRandom model represents a in-between case, since agents are uniformly distributed on the territory, but are still places within the Campania boundaries.

D-MASON Simulation. By noticing that most ABMs are inspired by natural models, where agents’ limited visibility allow to bound the range of interaction to a fixed range named agent’s Area of Interest (AOI), D-Mason adopt the so-called space-partitioning technique [15], where the agents’ world (the field) is split into tiles, each assigned to a LP.

Since citizens are basically moving on a map, our agents’ space consists in a rectangular area. Mason includes the Continuous2D field, where agents contained in it are located by a couple of continuous coordinates ranging from point (0,0) to point ([W]idth, [H]eight)). The distributed version embedded into D-Mason is called DContinuous2D: it retains all the features of the Continuous2D field, adding the support for two approaches to distribute the field and agents contained in it: dividing the space into vertical rectangles called rows (1-dimensional space partitioning or horizontal); or dividing it into a \(rows\,\times \,columns\) matrix (2-dimensional space partitioning or square, see Fig. 3). With our model, the square partitioning mode provides a significant speedup over the horizontal partitioning, lowering the communication effort while distributing the computational workload of the agents to LPs.

The behavior of agents is influenced by GIS data (map, zones and cities), nevertheless GIS data is static and does not require any synchronization among LPs. Reading ISTAT data, each LP manages an area of competence, and take care of agents that live in its area of competence. This is done by reading agent’s home location from the ISTAT dataset, looking for correspondent coordinates into GIS data, and converting it into 2-dimensional D-Mason coordinates.

3 Results

To evaluate the performances of the three models described above we developed the models in D-Mason. We have tested the models using two communications strategies available in D-Mason: AMQ (Apache ActiveMQ) that is the centralized communication strategy and MPI that uses the MPI standard [13, 14].

Simulations have been executed on a cluster of eight nodes, each equipped as follows: Hardware, CPUs 2 x Intel(R) Xeon(R) CPU E5-2680 @ 2.70GHz (#core 16, #threads 32) – RAM 256 GB; Software, Ubuntu 12.04.4 LTS – Java JDK 1.6.25 – OpenMPI 1.7.4.

3.1 Experiments Settings

We have investigated the scalability of the simulation considering the overall simulation time needed to simulate a 5 (simulated) minutes of real world system changing both the number of LPs and the simulation workload (# of agents). As described above the model uses a discretization time to simulate the real life clock. In our tests, the discretization time is 2400 steps per hour, the field size is \(3600\,\times \,2400\) and the neighbors’ influence radius is 1. D-Mason allows two kinds of space partitioning: 1-dimensional and 2-dimensional. After several pilot experiments, we have chosen the 2-dimensional partitioning due to the unbalanced nature of the positioning of the agents among the fields. The unbalanced density of agents will be a crucial part of our investigation. More details will appear in Sect. 3.2. Two kinds of experiments will be presented:

Simulate 5 min of Real Life. In this test we are interested in evaluating How much time is needed to simulate 5 min of real life? (which corresponds to 200 simulation steps). We tested four configurations which partition the field in \(4\,\times \,4\), \(6\,\times \,6\), \(8\,\times \,8\) and \(10\,\times \,10\) tiles assigned, respectively, to 16, 36, 64 and 100 LPs. Each configuration was performed on 2.5 million agents. Figure 1 shows the results for each model (Real, Random and CRandom). For each configuration, we show the total simulation time as well as how it is partitioned into the communication overhead (that includes the management overhead introduced by D-Mason) and the computation time.
Fig. 1.

Simulation performance with \(4\,{\times }\,4\), \(6 {\times } 6\) , \(8{\times } 8\), \(10\,{\times }\,10\) partitionings - 5 min of real clock.

The performance of the simulation is strongly influenced by the positioning model. The Random and CRandom models exhibit the same unimodal trend, as the number of LPs increase and manifest a balanced ratio of communication and computation.

The Real test provides the worst performance and unusual trend due to an unbalanced communication overhead. We investigated this problem analyzing the simulations with different agents positioning models and we discovered that this trend is due to the non uniform positioning of the agents (see Fig. 4).
Fig. 2.

Weak scalability. \(10\,{\times }\,10\) partitioning, 5 min of real clock.

Weak Scalability. This experiment aims to evaluate the simulation efficiency varying the total computation workload. We tested four configurations changing the total number of agents \(10\,\%, 40\,\%, 70\,\%\) and \(100\,\%\) (\(100\,\%=2.5\) mil). Each configuration was performed on a \(10\,\times \,10\) partitioning with 100 LPs. Figure 2 depicts the results of the three models using MPI as communication layer. Moreover we also compare the performances with the sequential version of the model implemented in Mason (we refers to this with the name SEQ). Random and CRandom tests provides a similar behavior showing good scalability. This results demonstrate the good performance of a 2-dimensional field partitioning on a uniform and quasi-uniform positioning density. On the other hand, the Real model manifests the worst scalability (just a bit better than the sequential version). This result is due to the communications overhead that is extremely irregular over the LPs. The Table 1 reports the speedup obtained during the weak scalability test. For each configuration, the minimum and maximum speedup are emphasized in bold. The best results are obtained by the Random model with 70% of computation amount and AMQ as communication layer; the worst performance is achieved by the Real positioning using the MPI communication layer. This confirms our hypothesis that the speedup is strongly related to agents positioning.
Table 1.

Experiments speedup varying the workload. \(10\,{\times }\,10\) partitioning, 5 min of real clock.

Workload

Test Name

10 %

40 %

70 %

100 %

AMQ - Real

3,36

2,49

1,97

1,74

MPI - Real

3,07

2,32

1,79

1,46

AMQ - Random

11,32

25,14

35,53

33,06

MPI - Random

7,63

20,01

27,29

27,78

AMQ - CRandom

7,69

21,13

29,14

31,86

MPI - CRandom

5,60

15,53

23,38

26,78

3.2 Analytical Analysis of ABM and GIS

Considering the results obtained in the preceding section, we decide to analytically evaluate the communication effort required by a GIS based distributed simulation that exploits a uniform 2D space partitioning approach (Fig. 3).

When a space partitioning approach is used, the amount of communication performed before each simulation step is related to: the size of the whole field (\([W]idth\,\times \,[H]eight\) in this specific analysis), the agents density distribution (d) i.e., the positioning of the agents over the field, the number of LPs (p), the maximum agents ride distance (\(\alpha \)), the maximum number of steps to perform a ride (\(\beta \)) and the agents area of interest radius (AOI) which depends on the neighbors’ influence radius (NIR).
Fig. 3.

D-MASON 2-dimensional uniform field partitioning on p tiles.

Recalling that using the space partitioning approach we have
$$\begin{aligned} AOI \ge \max \left( NIR ,\frac{\alpha }{\beta }\right) , \end{aligned}$$
(1)
therefore the AOI should be at least equal to the ratio \(\alpha /\beta \).
Since, in euclidean space, the diameter of our field is \(\alpha =\sqrt{W^2+H^2}\), one can easily verify that
$$\begin{aligned} {{W+H} \over 2} \le \alpha \le W+H. \end{aligned}$$
(2)
Hence, by using (1) and (2) we have that
$$\begin{aligned} AOI \ge \frac{\sqrt{W^2+H^2}}{\beta }\ge \frac{W+H}{2\beta } \end{aligned}$$
(3)
We can now evaluate the communication effort required by a GIS based distributed simulation. For each region, the communication effort \(\delta _c\) is obtained by counting the number of agents which belong to the edges of the region. The edges space, as shown in Fig. 3, is composed by 16 regions of sizes \({W \over \sqrt{p}}\,\times \,AOI\) (top and bottom), \({H \over \sqrt{p}}\,\times \,AOI\) (left and right) and \(AOI\times AOI\) (corners). The expected number of agents is obtained multiplying the size of the above described region by the density. Overall we have,
$$\begin{aligned} \delta _c= & {} p\,\times \,\left[ 4\left( {W \over \sqrt{p}} AOI\right) +4\left( {H \over \sqrt{p}} AOI\right) +8AOI^2 \right] \,\times \,d \nonumber \\= & {} 4\sqrt{p}\,\times \,d \times AOI \times (W+H) +8p\,\times \,d \times AOI^2 \nonumber \\\le & {} 8\sqrt{p}\,\times \,d \times \beta \times AOI^2 +8p\,\times \,d \times AOI^2 \nonumber \\= & {} 8 \sqrt{p}\times d\,\times \,AOI^2 \times (\beta + \sqrt{p}) \end{aligned}$$
(4)
where the inequality is due to Eq. (3).

Consequently the communication effort is linearly influenced by the AOI (which depend on the simulation model) and the density distribution of agents (d). In details the value of \(\delta _c\) varies according to the agents positioning over the field. When such value is irregular, the communication increases and affects all the regions since the whole system synchronizes before each simulation step.

This analysis motivates the poor performance of the simulation in the Real agents positioning experiment. Figure 4 depicts the positioning of the agents on the geographical zones in the Campania region. Real positioning provides a lots of zones with a small number of agents but there are also a small number of highly populated zones. Indeed, the density d over the field is non-uniform (the variance, in the number of agents per zone, is 302600129.2) and by Eq. (3) the communication effort \(\delta _c\) grows proportionally with the larger value of d.
Fig. 4.

Agents positioning over the region zones in Campania. In the figure is shown the frequency of zones in Campania with a certain density that ranging from 70 to 57869 people.

4 Conclusion

We considered the problem of simulation an ABM that uses GIS data. Exploiting GIS data in ABM is an important innovation in the ABM field. Several ABM examples [16] and users community [22] demonstrate the importance of this approach for improving the effectiveness of ABM model in complex systems study. Experimental results on a toy model, inspired by [2], demonstrate that the work partitioning, in a distributed GIS based ABM simulation is quite hard. According to our analysis, the main issue is the uneven positioning of the agents over the field, which jeopardize the performance of the simulation. Indeed, the speedup depends on communication effort \(\delta _c\) which, in a 2-dimensional uniform field partitioning approach, is: \(\delta _c \le 8 \sqrt{p}\times d\,\times \,AOI^2 \times (\beta + \sqrt{p}).\)

Therefore, the performance of the simulation scale up as a quadratic function of the AOI (which depends on the model) and is linearly influenced by the density distribution (d) of the agents over the field and the discretization time (\(\beta \)) used in the model. As future work, we plan to study the problem of balancing the communication effort among the LPs considering the density distribution of the agents over the field according to the field partitioning strategies.

References

  1. 1.
    Generative social science: studies in agent-based computational modeling (2006)Google Scholar
  2. 2.
    Andrighetto, G., Campennì, M., Conte, R., Paolucci, M.: On the immergence of norms: a normative agent architecture (2014)Google Scholar
  3. 3.
    Axelrod, R.: The complexity of cooperation (1997)Google Scholar
  4. 4.
    Balan, G.C., Cioffi-Revilla, C., Luke, S., Panait, L., Paus, S.: MASON: a Java multi-agent simulation library. In: Proceedings of the Agent 2003 Conference (2003)Google Scholar
  5. 5.
    Benenson, I.: Agent-based models of geographical systems. Int. J. Geograph. Inf. Sci. 27, 2483–2508 (2013)CrossRefGoogle Scholar
  6. 6.
    Brunsdon, C., Singleton, A.: Geocomputation: A Practical Primer. SAGE (2015)Google Scholar
  7. 7.
    Cioffi-Revilla, C.: Computational social science. Comput. Stat. Wiley Interdisc. Rev. 2, 259–271 (2010)CrossRefGoogle Scholar
  8. 8.
    Collier, N., North, M.: Repast HPC: a platform for large-scale agent-based modeling. In: Dubitzky, W., Kurowski, K., Schott, B. (eds.) Large-Scale Computing Techniques for Complex System Simulations. Wiley, Hoboken (2011)Google Scholar
  9. 9.
    Conte, R., Gilbert, N., Bonelli, G., Cioffi-Revilla, C., Deffuant, G., Kertesz, J., Loreto, V., Moat, S., Nadal, J.P., Sanchez, A., et al.: Manifesto of computational social science. Eur. Phys. J. Spec. Top. 214, 325–346 (2012)CrossRefGoogle Scholar
  10. 10.
    Conte, R., Paolucci, M.: On agent-based modeling and computational social science. Front. Psychol. 5, 668 (2014)CrossRefGoogle Scholar
  11. 11.
    Cordasco, G., De Chiara, R., Mancuso, A., Mazzeo, D., Scarano, V., Spagnuolo, C.: A framework for distributing agent-based simulations. In: Alexander, M., D’Ambra, P., Belloum, A., Bosilca, G., Cannataro, M., Danelutto, M., Martino, B., Gerndt, M., Jeannot, E., Namyst, R., Roman, J., Scott, S.L., Traff, J.L., Vallée, G., Weidendorfer, J. (eds.) Euro-Par 2011, Part I. LNCS, vol. 7155, pp. 460–470. Springer, Heidelberg (2012) CrossRefGoogle Scholar
  12. 12.
    Cordasco, G., De Chiara, R., Mancuso, M., Mazzeo, D., Scarano, V., Spagnuolo, C.: Bringing together efficiency and effectiveness in distributed simulations: the experience with D-MASON. SIMULATION Trans. Soc. Model. Simul. Int. 89, 1236–1526 (2013)CrossRefGoogle Scholar
  13. 13.
    Cordasco, G., Mancuso, A., Milone, F., Spagnuolo, C.: Communication strategies in distributed agent-based simulations: the experience with D-Mason. In: Mey, D., et al. (eds.) Euro-Par 2013. LNCS, vol. 8374, pp. 533–543. Springer, Heidelberg (2014) CrossRefGoogle Scholar
  14. 14.
    Cordasco, G., Milone, F., Spagnuolo, C., Vicidomini, L.: Exploiting D-Mason on parallel platforms: a novel communication strategy. In: Lopes, L., et al. (eds.) Euro-Par 2014, Part I. LNCS, vol. 8805, pp. 407–417. Springer, Heidelberg (2014) Google Scholar
  15. 15.
    Cosenza, B., Cordasco, G., De Chiara, R., Scarano, V.: Distributed load balancing for parallel agent-based simulations. In: 2011 19th Euromicro International Conference on Parallel, Distributed and Network-Based Processing (PDP) (2011)Google Scholar
  16. 16.
    Crooks, A., Wise, S.: GIS and agent-based models for humanitarian assistance. Comput. Environ. Urban Syst. 41, 100–111 (2013)CrossRefGoogle Scholar
  17. 17.
    D-MASON Official GitHub Repository. https://github.com/isislab-unisa/dmason. Accessed Jun 2015
  18. 18.
  19. 19.
    Federico, B., Squazzoni, F.: Complexity-Friendly Policy Modelling. Routledge, London (2015)Google Scholar
  20. 20.
    Gilbert, N., Troitzsch, K.: Simulation for the Social Scientist. McGraw-Hill, New York (2005)Google Scholar
  21. 21.
  22. 22.
    GIS, ABM community. http://www.gisagents.org/
  23. 23.
    Helbing, D., Balietti, S.: From social simulation to integrative system design. Eur. Phys. J. Spec. Top. 195, 69–100 (2011)CrossRefGoogle Scholar
  24. 24.
    Henrich, J., Ensminger, J., McElreath, R., Barr, A., Barrett, C., Bolyanatz, A., Cardenas, J.C., Gurven, M., Gwako, E., Henrich, N., et al.: Markets, religion, community size, and the evolution of fairness and punishment. Science 327, 1480–1484 (2010)CrossRefGoogle Scholar
  25. 25.
  26. 26.
    Kiran, M., Chin, L., Richmond, P., Holcombe, M., Worth, D., Greenough, C.: Flame: simulating large populations of agents on parallel hardware architecturesGoogle Scholar
  27. 27.
    Lazer, D., Pentland, A.S., Adamic, L., Aral, S., Barabasi, A.L., Brewer, D., Christakis, N., Contractor, N., Fowler, J., Gutmann, M., et al.: Life in the network: the coming age of computational social science. Science (New York, NY) 323, 721 (2009)CrossRefGoogle Scholar
  28. 28.
    Neumann, M.: Homo socionicus: a case study of simulation models of norms. J. Artif. Soc. Soc. Simul. 11, 6 (2008)Google Scholar
  29. 29.
    Paolucci, M., Kossman, D., Conte, R., Lukowicz, P., Argyrakis, P., Blandford, A., Bonelli, G., Anderson, S., de Freitas, S., Edmonds, B., et al.: Towards a living earth simulator. Eur. Phys. J. Spec. Top. 214, 77–108 (2012)CrossRefGoogle Scholar
  30. 30.
    Putterman, L.: Cooperation and punishment. Science 69, 1061–1075 (2010)MathSciNetGoogle Scholar
  31. 31.
    Regione Campania: Geoportale: Geographic information system of campania. http://sit.regione.campania.it/portal/portal/default/Cartografia
  32. 32.
    Rousset, A., Herrmann, B., Lang, C., Philippe, L.: A survey on parallel and distributed multi-agent systems. In: Lopes, L., et al. (eds.) Euro-Par 2014, Part I. LNCS, vol. 8805, pp. 371–382. Springer, Heidelberg (2014) Google Scholar
  33. 33.
    Schelling, T.C.: Dynamic models of segregation. J. Math. Soc. 1, 143–186 (1971)CrossRefGoogle Scholar
  34. 34.
    Squazzoni, F., Boero, R.: Complexity-friendly policy modelling. In: Innovation in Complex Social Systems (2010)Google Scholar
  35. 35.
    Sullivan, K., Coletti, M., Luke, S.: GeoMason: geospatial support for MASON. Technical report, Department of Computer Science, George Mason University (2010)Google Scholar
  36. 36.
    Tisue, S., Wilensky, U.: Netlogo: a simple environment for modeling complexity. In: International Conference on Complex Systems (2004)Google Scholar

Copyright information

© Springer International Publishing Switzerland 2015

Authors and Affiliations

  • Nicola Lettieri
    • 1
  • Carmine Spagnuolo
    • 2
    Email author
  • Luca Vicidomini
    • 2
  1. 1.ISFOLUniversità del SannioBeneventoItaly
  2. 2.Dipartimento di InformaticaUniversità degli Studi di SalernoFiscianoItaly

Personalised recommendations