[communications in computer and information science] trends in computer science, engineering and...

D. Nagamalai, E. Renault, M. Dhanushkodi (Eds.): CCSEIT 2011, CCIS 204, pp. 542–552, 2011. © Springer-Verlag Berlin Heidelberg 2011

Agent Based Modeling of Individual Voting Preferences with Social Influence

Vivek Kumar Singh, Swati Basak, and Neelam Modanwal

Department of Computer Science, Banaras Hindu University, Varanasi-221005, India

[email protected], [email protected], [email protected]

Abstract. Agent Based Modeling (ABM) is usually referred to as the third way of doing modeling & simulation in contrast to the historically popular equation-based macrosimulations and microsimulations. In ABM a system is modeled as a set of autonomous agents who interact with each other and also with the environment. The agents represent various actors in the system and the environment represents agent’s surroundings and the overall context. ABM has been successfully applied to model and analyze a number of complex social processes and structures, ranging from habitation patterns to spread of culture. In this paper, we present our experimental work on applying ABM to model individual voting preferences. The model explores process of formation of voting preferences, the factors governing it and its effect on the final voting pattern. We have incorporated social influence theory in our model and experimented with various settings. The paper concludes with a short discussion of the results and their relevance.

Keywords: Agent Based Modeling, Multi-agent Systems, Social Influence, Social Simulation, Voting Model.

1 Introduction

The development of social simulation over the past half century can be grouped into three broad periods: macrosimulation, microsimulation and agent-based models. Sociologists, particularly computational sociologists, first tried macrosimulations to model problems in supply-chain management, inventory control in a warehouse, urban traffic, spread of epidemics, migration and demographic patterns. Macrosimulations consist of a set of differential equations that take a holistic view of the system. However, taking the complete system as the unit of analysis had inherent limitations. Microsimulations focused on the use of individuals as the unit of analysis but continued with macro-level forecasting used in macrosimulations. Though microsimulations modeled changes to each element of the population but they did not permit individuals to directly interact or to adapt. The main focus remained forecasting macro effects of public policies that alter individual behaviours. ABM, the third methodology, takes a pure bottom-up approach and keeps the individual at the centre. It allows modeling individual actors in the population and their interactions.

Agent Based Modeling of Individual Voting Preferences with Social Influence 543

This makes it a highly suitable technique for analysis of emergent group behaviours resulting only from local interactions of individuals. Besides, there are various other advantages of this approach.

In Agent Based Modeling (ABM) a system is modeled as a set of autonomous agents, who can perceive the environment and act on the basis of some behavioural rules. The agents represent the actors in the system; environment represents the surroundings including neighbouring agents; and the behaviour rules model the interaction of the agent with other agents as well as with the environment. ABM can be used to model a number of phenomena in varied domains like Markets & Economy, Organizations, World Wide Web and Social Systems etc. Availability of fast and cheap computing power, coupled with other advances in computational sciences has paved the way for use of ABM as a preferred modeling and simulation technique. Since last few years ABM has become the frontrunner tool of the Sociologists and Psychologists who try to model social behaviours, particularly the behaviour of groups and societies. A large number of researches are now being carried out using this generative approach to model and analyze social phenomenon, such as spread of rumors, diffusion of ideas, emergence of cooperation, emergence of conventions & social norms, evolution and dynamics of spread of religions and cultures etc

This paper presents our experimental work on using ABM for understanding and analyzing voting preferences of individuals. We have modeled the process and factors shaping and affecting individual voting preferences for a particular political party. To keep the model simple and consequently analysis more accurate, we have taken individual voting preferences as ‘0’ and ‘1’ representing a two party political system. Every individual’s final voting decision is shaped and affected by a number of factors, such as his own perception about parties, his family views & beliefs, perception of his friends, and his neighbourhood preferences. Our simulation models the role and relative weights of these factors in shaping an individual’s voting preferences. We have also incorporated social influence theory in our model which states that similar persons are more likely to interact, and as a result of frequent interactions they are likely to become more similar. We experimented with multiple settings and configurations and obtained interesting results. The rest of the paper is organized as follows. Section 2 describes the ABM approach for modeling and simulation. Section 3 explains the process and factors believed to shape individual voting preferences. Section 4 presents the results of our experimental model. The paper concludes (section 5) with a short discussion of the results and their relevance.

2 Agent Based Modeling

Traditional modeling approaches to social systems relied on equation-based models that operate with a macro perspective. They operate on population attributes & their distributions and lack the focus on individual’s role. Equation-based models, though useful for macro-scale behaviour predictions, fail to model social systems (or processes) that lack central coordination, systems that are very complex in terms of their interdependencies and systems which produce novel emergent behaviours in

544 V.K. Singh, S. Basak, and N. Modanwal

absence of a clear understanding of the collective phenomenon. Taking this into consideration, Axtell [1] suggested that there are three distinct uses of Agent-based computational models of social systems: (a) when numerical realizations can be proposed and solved: agent models can be used as social simulations; (b) when a mathematical model can be formulated but can only be partially solved: agent based models can be a useful tool of analysis; and (c) when mathematical models are either apparently intractable or provably insoluble: agent based modeling is perhaps the only technique available for systematic analysis.

Although technically simple, ABM is conceptually deep. Modeling a complex social process requires high conceptual clarity and analytical ability. A key issue, therefore, is to decide when to use ABM for modeling social systems. An indicative list of situations when it is better to think and model in terms of agents is: (a) when there is a natural representation of actors as agents; (b) when the interactions between the agents are complex, non-linear, discontinuous, or discrete; (c) when agents exhibit complex adaptive behaviours; (d) when the population or topology of interactions is heterogeneous; and (e) when agents have spatial component to their behaviours & interactions. ABMs in social sciences involve human agents, whose behaviours are often complex, irrational and subjective. Therefore, one needs to think carefully about the social phenomenon at hand before going for ABM. Further, the model needs to be built at the right level of description, with only the right amount of details. Unnecessarily adding complexity to a model can make the analysis difficult and the model useless.

2.1 Steps in ABM

Designing an Agent-based model of a social system require first identifying the purpose of the model, i.e., the potential question to be answered. The typical steps to be followed afterward can be summarized as following: (a) identifying the agents and their behaviour rules; (b) identifying agent relationships and their interaction patterns; (c) selecting an ABM platform; (d) obtaining the required relevant data; (e) validating the agent behaviour model; (f) running the model and recording the outputs; (g) analyzing the outputs with a viewpoint of linking micro-scale behaviours of the agents to macro-scale behaviour of the system; and (h) validating the model outcomes and hence the model [2].

2.2 ABM Computational Resources

Many rich and easy to program software platforms and toolkits for ABM are now readily available. Some of the popular modeling tools (particularly for social sciences) are: Net Logo (http://ccl.northwestern.edu/netlogo/), Repast (http://repast.sourceforge.net/), Swarm (http://www.swarm.org) and MASON (http://cs.gmu.edu/~eclab/projects/mason/). Net Logo is a multi-agent programming language and modeling environment designed in Logo programming language. It is highly suitable for modeling and exploring emergent phenomena. Repast is basically an agent-based social network modeling toolkit but has


rich libraries to study dynamics of social processes as well. Swarm is a multi-agent simulation package to simulate the social and biological interactions of agents and their emergent collective behaviour. Swarm has two versions, namely Objective-C and Java versions. MASON is one of the latest Java platforms in this group. In addition to these platforms, there are many other toolkits & APIs available for modeling and visualizing social systems.

2.3 Benefits of ABM

Understanding social systems not only requires understanding the individuals that comprise the system but also how individuals interact with each other resulting in global behavioural patterns, which can be sufficiently novel. ABM helps researchers in investigating how large-scale macro effects arise from the micro-processes of interactions among many agents. Axelrod and Tesfatsion [3] call ABM the third way of doing science besides induction and deduction. ABM applied to social sciences take a methodical approach that could permit two important developments: (a) rigorous testing, refinement, and extension of existing theories that are difficult to evaluate using mathematical & statistical techniques; and (b) a deeper understanding of fundamental causal mechanisms in multi-agent systems. Bonabeau [4] goes a step further in proposing that ABM is a mindset and not merely a technology. He summarizes benefits of ABM over other modeling techniques in three statements: (a) ABM provides a natural description of a system; (b) ABM captures emergent phenomena; and (c) ABM is flexible.

2.4 Validating Agent Based Models

ABM, now an established scientific research practice, needs to incorporate a proper methodology of validation in order to verify the robustness of its findings and to truly act as a bridge between disciplines. Since simulation results of ABM are very sensitive to how agents are modeled, validation of computational model & simulation results is a critical issue in ABM [5]. Several approaches for validation of agent-based computational models have been proposed. Carley & Gasser [6] categorizes the validation approaches into three broad categories: (a) theoretical verification, which determines whether the model is an adequate conceptualization of the real world with the help of situation experts; (b) external validation, which determine whether the results from the model matches the results from the real world; and (c) cross-model validation, which compares the results of the model with other models. All these approaches are essentially aimed at validating the model at the macro level. Gilbert [7], on the other hand, emphasized that in order to completely validate a model it should be validated not only at macro-level but also at micro-level. Before going for a macro-level validation it is necessary to confirm that micro-level behaviours of agents are adequate representation of the actors in the system. Econometric validation and companion modeling techniques have been advocated to be suitable for empirical verification of social system models [8]. There are however still some problems in empirical validation of agent-based models [9], and validation continues to be one of the central problems of computer simulation methods including ABM [10].


2.5 ABM Applications in Social Science

ABM is currently being applied to model a variety of complex social phenomenon where simple local interactions generate emergent system-level behaviours. Some representative & relevant work can be found in [11], [12] & [13]. Macy & Willer [14] group most of the work on Agent-based modeling of collective behaviours in two categories: (a) models of emergent structure which includes works on cultural differentiation, homophilous clustering, idea diffusion, convergent behaviours and norms; and (b) models of emergent social order which include viability of trust, cooperation and collective action in absence of global control. Goldstone & Janssen [15] also identify three similar themes for Agent-based computational models of collective behaviour namely: (a) patterns and organizations which include settlement patterns & segregation, human group behaviours and traffic patterns; (b) social contagion which include spread of ideas, fashions, cultures & religions; and (c) cooperation which include evolution of cooperation, trust & reputations and social norms & conventions.

3 Modeling Individual Voting Preferences

Every individual while deciding his voting preference is directly or indirectly affected by a number of factors. The final voting decision of an individual can be taken as a function of these factors. An individual makes a preference about a political party based not only on his perception about the ideology & programs advanced by it, but by several other important factors. Individual may have an inherent preference formed over a substantial period of time from his perceptions. However, family, friends, neighbourhood (representing social and religious group membership) and mass media play an important role both in shaping an individual’s perception and deciding his final vote. We have modeled a population of individuals with their initial preference for a particular party, and then run the model by varying different factors to observe how (and possibly why) his initial favour changes to a particular final voting preference. Since the model deals with a two party system (used for simplicity of analysis), every individual is believed to have an initial favour for one of the two parties, represented as ‘0’ or ‘1’. Individuals during course of the time interact with friends, family members and neighbours. We have also associated feature sets with every individual to represent his overall mental and behavioural makeup. The friends of an individual are then defined as those having similar feature set values as that of the individual. Family of an individual represents the agents close to him and with whom he interacts very frequently. Family members are known to have a greater influence over an individual. Neighbours are taken as relatively distant agents on the grid and are viewed to represent individuals belonging to same social and religious groups of the concerned individual. An individual is thus believed to be affected not only by his friends and family members but also by the social and religious groups he may belong. Mass media also plays role in shaping an individual’s voting preference. We have devised three versions of the simulation: (i) model with social influence, (ii) a majority-based model, and (iii) model including friends, family and neighbours.


3.1 The Theory of Social Influence

The Social influence assumes that individuals (or agents) often imitate good behaviours through their interactions with other individuals. Its dynamics depends on familiarity of the interacting individuals, density of the neighbourhood, and popularity and spatial proximity of the other individuals. Interesting models of social influence have been proposed by Carley [16], Axelrod [17] and Axtell [18], Coleman [19], Nowak et al [20] etc. Axelrod, in his social influence model of dissemination of culture [21], focused on factors of local influence (tendency for people who interact to become more similar) and homophily (tendency to interact more frequently with similar agents). The more the agents interact, the more similar they become; and the more similar they become, the more likely they are to interact. Agents who are similar to each other interact with each other and become more similar. Axelrod’s model captures culture as a group of features with multiple trait values per feature. However, the emphasis is not on the content of a specific culture but on the way in which culture is likely to emerge and spread. The simulations with varying parameters regarding grid size, number of features and number of traits per feature resulted in polarization, despite the only mechanism for change being convergence towards a neighbour. We have used the basic idea of Axelrod’s culture model to represent friends of an individual. An individual’s friend is believed to have behavioural feature set similar to the individual. The feature set similarity represents their like mindedness of the individuals on various issues and hence possibility of having similar voting preferences or at least being affected by each other’s preferences.

3.2 Role of Family, Neighbourhood and Mass Media

An individual makes up his voting preference through a complex process involving non-linear feedback loop of interactions with a number of entities. Individuals interact more frequently with their family members and relatively less frequently with other persons belonging to their social and religious groups. An individual’s final voting value is no doubt a function of these factors. However, what is more important to understand is that what is the extent of this effect and the by what underlying process an individual finally makes up his voting preference. We have modeled this in our second and third simulation versions. The second version is a simple model to understand the dynamics of voting preference changes due to majority effect. In this model all individuals are assigned a random initial voting preference value and then his interactions with his 8-neighbour agents is observed and analyzed. The final voting preference is a stable value that results by a process of majority sampling. Since the initial voting preference values seem to play an important role on the final outcome, we experimented with multiple initializations. This also modeled the implicit favour (or political wave) introduced by the mass media. The third version of the simulation is relatively more complex and takes into account voting preferences of individual’s family and socio-cultural and religious neighbourhood. We have explored with different weights assigned to the influence of family and neighbours and observed their effect on the resultant complex set of interactions.


4 Experimental Setup and Results

We have designed our voting preference simulation using NetLogo [22]. The first model that we designed can be described as follows. A population of agents is created on a wrap-around agent grid of size 41 X 41. Every patch on the grid represents an individual. Every individual has certain attributes that characterize it. These attributes are represented as a feature set, which can be a 3 to 7 bit long string. Every bit in the string represents a particular trait of that attribute. The feature set values of two agents determine how similar they are. Two individuals will be similar if their feature set string contains common traits. For example, individual A (feature set: 00110) and B (feature set: 10110) are highly similar. In addition to the feature set, every agent also has a voting preference value, initialized randomly to either ‘0’ or ‘1’. Agents interact with various agents in their neighbourhood. According to social influence theory, it is more likely to interact and be affected by the agents similar to it. Interactions between similar agents make them more similar, represented by a change in their voting preference so that they get similar voting preference. The model selects an agent and samples its neighbourhood. Suppose n out of total m agents in the neighbourhood are similar to the agent. Then the voting preference of the agent changes to the majority preference of the similar agents (n). We obtained multiple runs of the model by varying both (i) the feature size, and (ii) initial voting preference value distribution. The feature size has been varied from 3 to 7 with feature similarity threshold being varied from 2 to 3 for feature set size of 5 and 3 to 7 for feature set size of 7. The initial voting preference assignments has also been varied from 10%-90% to 50%-50% for vote values of ‘0’ and ‘1’. Figures 1 and 2 show a snapshot of two runs of the model. With higher similarity threshold (in figure 1) there is less local polarization towards any vote value despite having initial favour for one vote value. Low feature set similarity however results in contiguous groups settling to same vote values.

Fig. 1. A snapshot of the run of feature set based simulation. The feature set size is 5 and similarity threshold is 3. In the figure on the left, the initial values of green = 757 and blue = 924 transform to final values of green = 730 and blue = 951. In the figure on the right, the initial values of green = 505 and blue = 1176 transform to final values of green = 417 and blue = 1264.


Fig. 2. A snapshot of the run of feature set based simulation. In the figure on the left, the feature set size is 7 and similarity threshold is 3 and the initial values of green = 505 and blue = 1176 transform to final values of green = 117 and blue = 1564. In the figure on the right, the feature set size is 7 and similarity threshold is 7 and the initial values of green = 169 and blue = 1512 transform to final values of green = 169 and blue = 1512.

In the second version of the simulation we did not use feature set concept, rather we chose to have a simple model where voting preference of an individual is simply taken as the majority vote value of its neighbours. For example, if more than 4 of the 8 neighbours of an agent prefer the vote ‘1’, then the individual is also likely to vote ‘1’ irrespective of his original preference. Hence the model is driven purely by the majority. At every tick, an agent sees in its neighbourhood and decides its own vote value based on the vote values of its 8-neighbours. Therefore, the majority vote value around an agent persuades him to align with it. In case of tie, that is in the 8-neighbourhood 4 agents have ‘1’ vote value and the remaining 4 have ‘0’ vote value, the agent’s original vote value prevails. Since the final voting value of an individual depends only on the majority voting value present in the 8-neighbourhood and that the initial vote values are set randomly; we thought it would be interesting to see what happens when the initial vote values of agents have a non-uniform distribution. We run the model with multiple initial vote value distributions. The initial vote value assignments have been varied from 10% agents having ‘0’ vote and remaining 90% having ‘1’ vote, to many other combinations such as 20%-80%, 30%-70%, 40%-60%, 45%-55% and 50%-50% of ‘0’ and ‘1’ votes. A snapshot of the two interesting runs is shown in figure 3. The results demonstrate that the final voting preference of the population in the majority model depend heavily on the initial distribution of voting preferences. Whenever there is a favour (about 60% against 40%) for initial distribution of one vote value over the other (say ‘0’ over ‘1’), then the final voting preference of most of the individuals becomes the favoured one (‘0’ vote). In case of close to equal distribution of both the vote values (‘0’ and ‘1’ being less than 10% different), the final voting preference of the majority may tilt relatively favourably on either of the sides.


Fig. 3. A snapshot of the run of majority based simulation. In the figure on the left, the initial values of green = 1561 and blue = 1040 transform to final values of green = 2345 and blue = 256. In the figure on the right, the initial values of green = 1171 and blue = 1430 transform to final values of green = 659 and blue = 1942.

Fig. 4. A snapshot of the run of modified majority simulation. In the figure on the left, the initial values green = 505, blue = 1176 & weights own-vote = 80%, family-vote = 15% and neighbour-vote = 5%; results into final values green = 413, blue = 1268. In the figure on the right, the initial values green = 841, blue = 840 & weights own-vote = 40%, family-vote = 40% and neighbour-vote = 20%; results into final values green = 843, blue = 838.

The third version of the model that we explored is a modified majority model. In this version an individual’s voting preference is determined not simply by the prevalent majority vote, but trough a weighted average of values of his own vote, his family members’ vote and his neighbours’ vote. We chose 4-neighbourhood to represent family and 8-neighbourhood to represent neighbours. Here, 8-neighbours stand to represent members of individual’s socio-cultural and religious groups. Since


the 8-neighbourhood also contains the 4-neighbourhood on the grid, we have chosen relatively smaller weight for neighbours. We have run the model by varying the weights in weighted average computation and also the initial vote value distribution. Figure 4 presents the results of two such runs. The results demonstrate that despite giving higher weight to the individual voting preference, the initial bias introduced in the population results to a certain degree of local polarization. Only very high value of own vote is able to sustain the other voting preference. In case of approximately equal initial voting preference also, high own-vote weight value tend to preserve most of the initial voting pattern. As the weight assignments tend to become relatively more in favour of the family and neighbourhood combine, local polarization is observed. The results in this case become somewhat similar to the majority-based version of the simulation.

5 Conclusion

The ABM methodology is a favorable approach to model and analyze complex social phenomena that may involve non-linear feedback loops. It has been applied successfully to model a number of social phenomena involving different social processes and organizational structures. In this paper, we have demonstrated how ABM can be applied to model and analyze the voting preference formation and resultant voting decisions of individuals in a population. The model assumes a two party system. We designed three versions of the simulation and observed the results for a large number of runs with different parameter variations. The initial voting preference distribution in the population, role of friends (likeminded individuals), affect of family members and neighbours; all produce interesting results to analyze. Social influence theory incorporated into model shows that individuals’ voting decisions are affected more by those who are similar to them. However, if the similarity threshold is relatively lower, the model results in local voting preference polarizations. When the model is based only on majority, the initial distribution of voting preferences becomes the key deciding factor. These results taken together resemble to actual voting patterns observed in closely knit societies and segregations such as far off villages and tribal groups. The role of implicit favorable waves introduced by mass media can also be understood by the results of the second version of the simulation. When the weights of factors playing role in shaping an individual’s voting preference are varied, the results demonstrate that high weight to individual vote is able to sustain variations in voting preferences. When the own-weight is reduced to give more favour to the family and neighbours, polarizing tendency is seen. Voting behaviour and patterns of relatively informed and well-educated persons conform to this trend. The experimental simulations attempt to model an interesting social phenomenon and present interesting results that need to be further analyzed to draw relevant inferences, analogies and explanations.

References

1. Axtell, R.: Why agents? On the Varied Motivations for AgentCcomputing in the Social Sciences - working paper 17. Centre on Social and Economic Dynamics, Brookings Institution, Washington D.C (2000)


2. Macal, C.M., North, M.J.: Tutorial on Agent Based Modeling and Simulation. In: Kuhl, M.E., Steiger, N.M., Armstrong, F.B., Joines, J.A., Young, M. (eds.) Proceedings of the 2005 Winter Simulation Conference, The Technical Writer’s Handbook, University Science, Mill Valley (1989)

3. Axelrod, R., Tesfatison, L.: A Guide for Newcomers to Agent Based Modeling in the Social Sciences. In: The Handbook of Computational Economics. Agent Based Computational Economics, vol. 2 (2005)

4. Bonabeau, E.: Agent-based Modeling: Methods and Techniques for Simulating Human Systems. Proceedings of National Academy of Science 99(3), 7280–7287 (2001)

5. Takadama, K., Kawai, T., Koyama, Y.: Micro- and Macro- level Validation in Agent-Based Simulation: Reproduction of Human-Like Behaviours and Thinking in Sequential Bargaining Game. Journal of Artificial Societies and Social Simulation 11(2), 9 (2008)

6. Carley, K.M., Gasser, L.: Computational and Organization Theory. In: Weiss, G. (ed.) Multiagent Systems - Modern Approach to Distributed Artificial Intelligence, pp. 299–330. The MIT Press, Cambridge (1999)

7. Gilbert, N.: Open Problems in Using Agent-based Models in Industrial and Labor Dynamics. In: Leombruni, R., Richiardi, M. (eds.) Industry and Labor Dynamics: the Agent Based Computational Approach, pp. 401–405. World Scientific, Singapore (2004)

8. Moss, S.: Alternative Approcahes to the Emperical Validation of Agent- Based Models. Journal of Artificial Societies and Social Simulation 11(1), 5 (2008)

9. Windrum, P., Fagiolo, G., Moneta, A.: Emperical Validation of Agent-Based Models: Alternatives and Prospects. Journal of Artificial Societies and Social Simulation 10(2), 8 (2007)

10. Kuppers, G., Lenhard, J.: Validation of Simulation: Patterns in the Social and Natural Sciences. Journal of Artificial Societies and Social Simulation 8(4) (2005)

11. Epstein, J.M.: Generative Social Sciences: Studies in Agent-based Computational Modeling. Princeton University Press, Princeton (2007)

12. Gilbert, N., Troitzsch, K.: Simulation for the Social Scientist. Open Univ. Press, Buckingham (1999)

13. Epstein, J.M., Axtell, R.: Growing Artificial Societies: Socail Science from bottom-up. MIT Press, Cambridge (1996)

14. Macy, M.W., Willer, R.: From Factors to Actors: Computational Sociology and Agent Based Modeling. Annual Review of Sociology 28, 143–166 (2002)

15. Goldstone, R.L., Janssen, M.A.: Computational Models of Collective Behaviour. Trends in Cognitive Sciences 9(9) (2005)

16. Carley, K.M.: A Theory of Group Stability. American Sociological Review 56, 331–354 (1991)

17. Axelrod, R.: The Complexity of cooperation. Princeton University Press, Princeton (1997) 18. Axtell, R., Axelrod, R., Epstein, J.M., Cohen, M.D.: Aligning Simulation Models: a case study

and results. Computational and Mathematical Organization Theory 1(2), 123–141 (1996) 19. Coleman, J.S.: The Use of Electronic Computers in the Study of Social Organizations.

European Journal of Sociology 6, 89–107 (1965) 20. Nowak, A., Jacek, S., Latane, B.: From Private Attitude to Public Opinion: A dynamic

theory of social impact. Psychological Review 97, 362–376 (1990) 21. Axelrod, R.: The Dissemination of Culture: A model with local convergence and global

polarization. Journal of Conflict Resolution 41, 203–226 (1997) 22. Wilensky, U.: NetLogo. Centre for Connected Learning and Computer-Based Modeling,

Northwestern University, Evanston, IL (1999), http://ccl.northwestern.edu/netlogo

[communications in computer and information science] trends in computer science, engineering and...

Documents