Eighth Annual Swarm Users/Researchers Meeting SwarmFest 2004  http://cscs.umich.edu/swarmfest04/ The Center for the Study of Complex Systems University of Michigan Ann Arbor, Michigan USA May 9 - 11, 2004

Poster Abstracts (First author's name in bold) Extended Abstracts, Full Papers and Slides, when provided, are available as links off the abstract title   Xdrone: Parallel Batch Runs of Agent-Based Models ... For the Rest of Us Ed Baskerville ebaskerv@umich.edu Center for the Study of Complex Systems, University of Michigan       Xgrid, a recent technology initiative from Apple, enables researchers to easily group networked Mac OS X computers into a parallel computational grid. Xdrone, a plug-in for Xgrid, brings this technology to batch runs of agent-based models. With Xdrone, you can easily harness any network of Macs to explore your model's parameter space—even, for example, an ad-hoc wireless network. Once you set up parameters and begin the batch, runs are automatically assigned to computers on the grid as they become available.       Xdrone is modeled after Ted Belding's Drone tool in both name and design, and models designed for Drone will work with Xdrone unmodified. In addition to support for text configuration files, Xdrone provides a straightforward graphical interface for setting model parameters and performing batch runs. Xdrone can collect data to a local hard disk or to a network directory. Furthermore, if you add one additional command-line option to your model program, Xdrone can query the program to detect what parameters are available and modify the user interface accordingly.       Xdrone makes one aspect of agent-based modeling a little easier to manage. By itself, this is unremarkable—performing batch runs is hardly the most difficult part of ABM. Still, the wider trend toward easier design, implementation, and management of agent-based models is an important one. Alongside many other developments, improved ease of use will help promote ABM's wider acceptance. An Agent-Based Model of Human Achievement and Self-Efficacy Development Paul Chiusano1 Alex Chovanec1 Mike Samples1 1{pchiusan, achovane, msamples}@umich.edu Center for the Study of Complex Systems, University of Michigan.       The notion that most people are not taking close to full advantage of their abilities or potential to achieve is a widespread cultural sentiment. While "overachieving" is arguably a stabler state than underachieving, underachievers still not only exist in society--they appear to constitute the majority of individuals. Similarly, while variety is the "spice of life," it seems that many individuals opt to confine their life efforts to a small number of endeavors. We are interested in developing a model of human achievement which captures some of the rich, complex behavior observed in individuals and groups of individuals developing their abilities and self-efficacy.       Our approach differs from adaptive specialization models in organizational theory and insect sociology in that we are explicitly modeling agents' beliefs. We consider a model of human achievement in which agents are unaware of both their own abilities and the actual difficulty levels of activities in the world and rely instead on their (not necessarily accurate) estimations of these values. We draw on the extensive literature in social cognitive theory on motivation and self-efficacy to guide us toward reasonable agent learning rules.       The world for our model consists of a collection of agents connected by some social network. The world has various types of activities that need to be (or can be) performed by agents. Activities may be competitive (like publishing a paper) or uncompetitive (whistling). Activities have an actual difficulty level (ADL) associated with them.       Agents have parameters associated with each type of activity in the world: actual ability and perceived ability. They also have a function which evaluates the activity's perceived difficulty. The model is run by having a "fate" function continually select agents for activities (alternately, agents may volunteer for activities). An agent can choose to participate, and his decision is based on his expected likelihood of success (perhaps perceived ability - perceived difficulty).       Perceived difficulty is based not only on the agent's own experiences and attitudes, but on those of his neighbors. Thus, agents learn vicariously: when they observe those around them succeeding at an activity, they decrease their perception of the activity's difficulty and become more likely to attempt it themselves. Likewise, when agents observe those around them failing, the agents' perceptions of difficulty increase, and they become less likely to attempt the activity. As is suggested by social cognitive theory, agents are most influenced by others that seem similar to themselves.       If an agent decides to participate in an activity, we examine his actual ability and probabilistically return a "success" signal if it is greater than the ADL of the activity. Otherwise, we send a "failure" signal. In either case, the agent's actual ability is increased and the agent incorporates whatever feedback he receives. This in turn affects his future decisions (an agent who was successful at an activity in the past is more likely to attempt it in the future, etc).       Running the model, we might hope for the emergence of regions of overachievers and underachievers, or, in the case of a multi-activity world, the emergence of specialization. We can then ask questions like: what traits characterize successful vs. unsuccessful agents? What are the most "effective" groupings of agents and strategies for success? Do agents usually have accurate perceptions of their abilities? There is a connection here to attribution theory--we might ask which explanatory styles (if any) lead to convergence of an agent's perceptions with reality. Which explanatory styles lead to greatest success for an agent? (Paper in pdf) Dissemination of Culture using a Quantum Model Scott Christley schristl@nd.edu University of Notre Dame       Axelrod's cultural dissemination model introduces an agent-based simulation where random agent interactions transmit culture through an agent population, and the system evolves over time to form multiple stable homogeneous cultural regions. We expand upon this work by introducing a quantum model. Agents are represented by quantum registers, and agent interactions are quantum operations performed on those registers. Results indicate that multiple stable hetergeneous cultural regions form, the number of regions is greater in the quantum model, and there is a greater diversity in the sizes of the cultural regions. How information Security Key Challenges can be Faced Using Multi-Agents Based Technologies Fabio Ghioni fabio.ghioni@telecomitalia.it Telecom Italia Group       When we refer to Information Security, we are speaking about a complex environment. Actually, the combination of people, networks and IT systems interacting creates a degree of infrastructure's complexity for which is hard to achieve and maintain security standards and procedures via the traditional approach.       Complex systems have so many variables and interacting forces that the traditional, linear approach is not working. A suitable emerging alternative is represented by the multi-agents approach, where many software agents interact with objects and other agents in an adaptive way.       1. Agents characteristics and Information Security A standard definition of agents and multi-agents does not exists, but they are playing an important role in artificial intelligence due to their ability to take initiative, to communicate and have certain responsibility. In some cases, their able to learn from experience. A growing number of applications, even military ones, are based on agents but a lot more has to be done to exploit agents and multi-agents possibilities in the Information Security field. This kind of application is increasing in importance as pervasive mobile computing become more present in our everyday life joint with a greater complexity of the whole system. Due to their peculiarities, multi-agents based technologies can be profitably employed in management and optimisation of information security key problems in complex environment.       1.1 Information classification and dynamic environment A key topic to confront with, when approaching information security, is the correct information classification, in order to avoid the disclosure of sensitive data to unauthorized people while giving to authorized personnel the correct and timely data they need. Due to the highly dynamic nature of business, to the great and dispersed amount of documents and data, and to their very short life cycle, a manual categorization and classification is a highly error prone process and does not stand as a suitable solution. To load the burden, a great number of processes, applications, and personnel are accessing such information in different ways to achieve different goals. In such a scenario the multi-agents contribution in information retrieval and classification can be definitely the right answer for the automation and optimisation of this classification task.      1.2 Overall system security framework and agents The other Information Security key field that could benefit of multi-agents approach is the system and network protection from attacks. The growing complexity of networked IT systems, comprising different devices, operating systems, protocols and languages bring to attention the need for a security framework, with self diagnostic ability and efficient incident handling procedures. Multi-agents are able to gather log information from different devices and format, then decide if the pattern registered can be a malicious unauthorized activity or a casual accident. On this basis, only the significant security related incidents can be brought to human operator attention, thus increasing his ability to manage security threat because he handle data really significant. Fruit and Meat: growing trading among wild prehistoric humans guys Gianluigi Ferraris ferraris@econ.unito.it Affiliation?       This work tries to find a plausible motivation for the emergence of the commerce into a prehistoric proto society, uniquely based upon economics. The starting idea is that exchanging goods implies a greater benefit, from a society "as a whole" standpoint, than fighting for their possession. No ethical matters have been kept in mind: to steal goods is accepted behaviour as much as to trade. By means of a Swarm model the interaction among human agents and between the humans and the environment has been simulated. The obtained results reveal that trading tends to replace the fighting until the whole society becomes based on trading, even if the starting population was composed by am overwhelming majority of fighters. While completing the present version and making it consistent, further efforts will be spent enriching it, in order to study more complex topics as, for instance, the emergence of simple institutions and the like. (Presentation in ppt) Next Generation Models of Farm Management and Rural Change Tyler Freeman trf486@mail.usask.ca Dept. of Agricultural Economics, University of Saskatchewan       Agricultural modeling, as an applied science, is driven by two general concerns. The first is the problem of building a thorough understanding of the system under analysis and, secondly, to predict potential changes to that system. Agriculture is a complex system of individual farms operating within an equally complex and dynamic environment. The importance of understanding the interaction between individual farm operators is particularly evident in the competition for limited land resources. In general, current models of farm behavior and investment, do not adequately account for the interactions between individual farms and the spatial region within which they operate. As a result, these models have a limited ability to predict the long term structure of individual farms and rural regions. Rural societies are significantly impacted by shifts in farm structure through the influence producers have on local markets, demand for education and health services and other rural infrastructure. Agent-based models promise to overcome the limitations of existing farm-level models and will allow researchers the ability to better understand the dynamics of a rural region which are driven primarily by individual farm management decisions. A basic agent-based model of a western Canadian agricultural region, limited to annual crop production, is developed on the NetLogo platform to examine the impacts of producer risk attitudes and government policy on farm and rural structure. The primary objective of the paper is to evaluate the role of agent-based modeling as a tool for appraising and predicting changes in farm and rural structure, specifically related to farm size and management practices. Modeling Sustained Evolution in RePast Laszlo Gulyas gulyas@sztaki.hu Computer and Automation Research Institute, Hungarian Academy of Sciences, Budapest, Hungary George Kampis kampis@hps.elte.hu History and Philosophy of Sciences, Eotvos University, Budapest       Can we produce an existence-proof model, akin to von Neumann's model of self-reproduction, that exhibits open-ended evolution, with increasing diversity and complexity? This was the first challenge for agent-based modeling on John Holland's list addressing the audience of SwarmFest in 2003. Evolutionary algorithms, as known today, converge to an externally defined optimum and settle at a moderate level of diversity.       In this presentation we deal with the problem of open-ended evolution of new species, and prove that agent-based modeling techniques are capable of tackling this problem in ways inapproachable by traditional equation-based modeling. We present an agent-based simulation, implemented in RePast, that is based on the idea of 'fat' phenotypes which generate a changing interaction space. Such changes can, in turn, define new selection forces. In our sexual selection-based model, species are defined as reproductively isolated and functionally different sub-populations. New species occur when genetic mutations produce individuals with a new phenotype that lets them disregard existing mating preferences and thus redefines previously relevant interactions. Constructing an Agent-Based Model of the Spread of Tuberculosis Kristen Hassmiller khassmil@umich.edu Department of Health Management and Policy, School of Public Health       Traditional models of the spread of disease assume perfect mixing. This implies that every individual is equally likely to infect any other individual. However this assumption is far from realistic. Agent-based modeling permits investigation of how different epidemics look when the social networks tying individuals together differ.       For this poster presentation, I consider the specific case of tuberculosis. I will present preliminary findings on the spread of tuberculosis through a simple simulated population with different forms of underlying social networks. Based on work by Watts and Strogatz (Nature, 1998), I will consider the spectrum from regular networks, to small world networks, to random networks, comparing these to the ABM approximation of the traditional mean-field ordinary differential equation model. I will also consider the generation of social networks based on rules of interaction (i.e. local employment patterns, educational and transportation systems, and military service) such as that used in Epstein's ABM smallpox model (Brookings Institution Press, 2004).       I will also discuss methodological issues in making the simple tuberculosis model more realistic, including: how to model birth and death; updating social networks over time; incorporating heterogeneity of agents (which impacts agents' risk of infection, progression to active disease, time between active disease and diagnosis, adherence to treatment); and the trade-off between simplicity and realism in the model. The City Development Model Based on Multi-Agents Fang Jing Jingfangfx22@yahoo.com.cn Department of Computer Science and Engineering, Nagoya Institute of Technology, Japan Prof. Desheng Du Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Prof. Takahashi Nagoya Institute of Technology, Japan       With expanding the CA(Cellular Automata) model, we adopt multi-agents technology to establish the city development model. In consideration of many layers structure and complexity of the city development model, each of the city function bureaus of the main economic characters that affect the resident was mapped by CA in many layers structure. The main frame of city development model was built by the agent.s migration behaviors that represent the interaction of the resident on each CA. The artificial experiment on the development of Olympics city shows the actual possibility and theories' value of the method. (Paper in pdf) Bid Quote Price Game Model Prof. Xu Jing jing_xu45626@163.com Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Jiapeng Helian Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Xiaobo Sun Department of Automation, Computer and Control College Haerbin University of Science and Technology, China       In this paper the bidder multi-agent game model was investigated. The author look on the bidder's quoted price using the idea of dynamic game based on the first price sealed game theory. The primary factor whether the bidder will succeed or not is the bidder agent's cognitive ability in the game in the case of the fixed ability. The paper employs the computational model of the single layer perceptron and the XOR function to get the math map then we can get the description of the cognitive ability in the dynamic game. The competitive-bidding model was simulated in the swarm flat. The Artificial experiment shows that the model has the rationality and the further research value. (Paper in pdf) Modeling in Battle Damage Based on Multi-agent Prof. Xu Jing jing_xu45626@163.com Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Liying Yong Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Dr. Hongping Pan Department of Automation, Computer and Control College Haerbin University of Science and Technology, China Xiaobo Sun Department of Automation, Computer and Control College Haerbin University of Science and Technology, China       The establishment of traditional model about battle damage is static by use of the statistics from the real battlefield without considering the actual battle environment well. However, complex adaptive system theory is another new alternative for research on dynamic battle damage. In this paper, we attempt to establish a simple dynamic battle damage model with a dynamic battle environment on Swarm in order to illustrate the methodologies of modeling based on the fact and reasonable presumption. (Paper in pdf) ART (Artificial Reasoning Toolkit) Library Marco Lamieri lamieri@econ.unito.it University of Turin, Italy Gianluigi Ferraris University of Turin, Italy       The ART (Artificial Reasoning Toolkit) is a pure Java library devoted to handle Genetic Algorithms and Classifier Systems.       It has been engineered in order to be used into Swarm or others agent based simulation's models, to easy obtain minded agents who are fully autonomous, able to decide their own behaviors and able to change it to fit in different environmental conditions. Another main usage of the algorithm is to search bounded optimal solutions in very wide solution spaces and for quite undefined problems. This kind of problems are solved using the convergence method: the best result is assumed to be achieved when a given convergence of the same solution exist in the population. It is widely accepted as mathematical proof that the genetic algorithm, due to its fitness-proportionate reproduction, converges to better solutions.       The genetic algorithm's implementation, starting from John Holland's work, introduces some extensions and innovations: extended alphabet: each gene can be represented by up to 32000 values. In a standard representation the genes have a binary alphabet and so the genomes have to be explicitly translated into the various aspects composing the solution, which after some manipulation, as crossover or mutation, can become meaningless. With the extended alphabet each allele can be a meaningful part of the solution and the translation process is easier.       multi genome: each individual of the population is represented by a chromosome that could be composed by a variable number of genomes. Each genome of a chromosome represent a substrategy and the chromosome is the genetic algorithm's formalism for a strategy driving the actions of the simulated agent. The multi genome schema give a high degree of freedom to the user in formalizing problems in which coexist different binded aspects.       rescale fitness operator: the natural selection process has been modified in order to improve efficiency and manage negative fitness values. The technique utilized consist in rescale the fitness of all the chromosome. univocal genome: using this option each value of the alphabet is unique within the genome, it means that in a genome there can not be two or more identical genes. (poster in ppt) MASON: A New Multi-Agent Simulation Toolkit Sean Luke sean@cs.gmu.edu Claudia Cioffi-Revilla, Liviu Panait, Keith Sullivan. Dept. of Computer Science and Center for Social Complexity George Mason University       We introduce MASON, a fast, easily extendable, discreteevent multi-agent simulation toolkit in Java. MASON was designed to serve as the basis for a wide range of multiagent simulation tasks ranging from swarm robotics to machine learning to social complexity environments. MASON carefully delineates between model and visualization, allowing models to be dynamically detached from or attached to visualizers, and to change platforms mid-run. We describe the MASON system, its motivation, and its basic architectural design. We then discuss five applications of MASON we have built over the past year to suggest its breadth of utility. (paper in pdf) Simulating an Artificial Society: The Free/Open Source Software Community Greg Madey gmadey@nd.edu Computer Science & Engineering, University of Notre Dame       We report the latest results from an ongoing study of Free/Open Source Software (F/OSS) development at the community level. A computer simulation of the F/OSS community is developed using Java/Swarm and a relational database. Empirical data is used to parameterize the simulation, which in turn is used to investigate hypotheses about processes and mechanisms leading to F/OSS community formation. Successive computer experiments using the simulation have been conducted to test various hypothesis about mechanisms and processes at play in the F/OSS developer community. Publicly available data about F/OSS projects, developers, processes, and their relationships have been collected from F/OSS hosting sites, including SourceForge and others. Numerous descriptive statistics, including the existence of many power-law relationships, are presented. The F/OSS community is modeled as a collection of ad hoc, social networks consisting of heterogeneous agents, self-organizing into projects and clusters of projects. The quantitative data, the model, and the simulation offer insight into F/OSS project coordination and communication. Web-Based Molecular Simulation using Agent-Based Modeling Techniques Greg Madey gmadey@nd.edu Computer Science & Engineering, University of Notre Dame       Note: This poster reports on an ongoing study that began at SwarmFest 2002 (Seattle, WA) when one of the investigators presented a research problem and requested opinions on the feasibility of using Swarm on the study. Those opinions, along with practical advice, contributed to the following successful results.       Natural organic matter (NOM), a heterogeneous mixture of molecules, plays a crucial role in the evolution of soils, the transport of pollutants, and the change of global weather. The evolution of NOM over time is an important research area in biology, geochemistry, ecology, soil science, and water resources. Due to its complexity and structural heterogeneity, new simulation approaches are needed to help to better understand the structure and the evolution of NOM. We present a new stochastic model, implemented using Java/Swarm, which explicitly treats NOM as a large number of discrete heterogeneous molecules. The NOM, micro-organisms, and their environment are taken together as a complex system, and simulated using an agent-based modeling approach. The global properties of NOM evolution over time can be studied by simulating the physical and chemical reactions between individual agents with temporal and spatial properties. Unlike the previous stand-alone simulation models, the NOM simulation model serves as an example of E-science, in which we do science on the Web by combining recent information technologies (Java 2 Enterprise Edition, J2EE) with an agent-based computational approach. An intelligent Web-based interface is developed to allow scientists to access the remote simulation model from a standard Web browser. The Web-based interface enables scientists to remotely provide parameters for their simulations, start and stop the simulations, and view the results. The initial users of the NOM simulation model includes a geographically separated group of NSF sponsored scientists from different research areas. A NOM collaboratory is built to promote collaboration among these scientists and allow them to share their data and information across distributed sites. A XML-based Markup Language, NOML, is provided to build the XML-based Web components and facilitate Web services development in the future. Penelope Meets Nemote: Distributed Production Planning Optimization Matteo Morini matteo.morini@unito.it University of Turin       The original textile-oriented production planner Penelope is an extremely timeconsuming process, very heavy from a computational standpoint. Running the system as a monolythic task, on a single CPU, brings unsustainably long completion times, especially when employed in production environments where the size of the planning problem exceeds the naivet of unsophisticated situations tailored for testing needs only.       In order to overcome prompt response constraints, the inherently parallel work of evaluating multiple candidate plans has been distributed among multiple CPUs, residing on networked pcs.       The tasks distribution, load balancing and failure tolerance management is performed by an infrastructure developed by the Nemote* group: Riccardo Boero, Gianluigi Ferraris, Matteo Morini and Michele Sonnessa. The original, swarm-based, objective-C model has been split into self-contained components glued together by java processes. Different nodes comminicate via RMI and the java and objective C parts communicate via tcp sockets.       Performance scales almost linearly, thanks to the careful trimming of the system, which allows tasks to be distributed in batch in order to minimize network overhead.       Nemote has not been developed as an ad-hoc tool, being easy to exploit whenever distributed computational needs arise, when parallel modelization is the fundamental issue, when remote interaction among cooperative distributed processing leads to more plausible simulations.       (*) NEtworked MOdelling TEam, supported by the Liases Computer Lab of the Faculty of Economics, Using Swarm to Model Iterated Language Development Games Meredith L. Patterson University of Iowa Robert Arens University of Iowa Tristan Thiede University of Iowa Robert J. Hansen University of Iowa {mlpatter, rarens, tthiede, rjhansen}@cs.uiowa.edu       Recent work by Dr. Teresa Satterfield has used Swarm to model creolization, the development of a new language which bridges two unrelated parent languages. We extend her approach to include a game-theoretic model of lexical acquisition, morphological acquisition, and shifts in social status. Within the Swarm framework, agents interact as speakers and listeners in a noncooperative bimatrix game whose payoffs represent a tradeoff between language acquisition and individual utility value. Each interaction between agents involves one or more iterations of the game, using mixed strategies which can vary situationally, and over the course of many iterations over the entire swarm, the change in agents' lexical and morphological inventories reflects the development of a creole. Change in social status is represented as a lottery among agents who have acquired a certain degree of language competence, requiring agents to employ delayed-gratification strategies in order to achieve the higher utility payouts granted by higher social status. The game and lottery representations allow for a finer degree of control over situational variables, and will facilitate further work in modelling other elements of synchronic and diachronic language change, e.g. sound change and syntactic parameter shift. Integration of Multiple Temporal, Spatial Scale Processes in a Common Modelling Framework. Praveena Pepalla praveena@cc.usu.edu Utah State University Paul Box paul.box@usu.edu Utah State University       Modelling a framework for an independent system helps in understanding the behaviour of that particular system. It is rare that systems exist independently with out interaction with other systems. In a dynamic landscape, what we observe as global behavior is really the interaction of different systems like physical, hydrological and social systems, which are defined at a variety of temporal and spatial scales; the definition of these subsystems can make them incompatible for integration into a global framework. Some of these systems are spatially static. Others change their locations at every time step exchanging feedbacks present at that time and location. Since all of these systems can be conceived as agents, agent based simulation using swarm protocals can be used as an integrated platform to predict the overall impacts of mobile agents on the environment and vice versa. This model can also be used to predict the effects of inividual agents and also the aggregate of two or more agents. An example is presented for a study in the Luquillo rainforest of Puerto Rico, where Cellular Automata is used to model landscape and hydrological processes, and free-roaming agents are used to depict shrimp migrations and recreational users in the forests river systems. Integration of the various subsystems is implemented through swarms list management structure, allowing a global world to be assembled at any time step according to the individual perceptions of any participant, be they a shrimp, a person, a group of people, a pool, or a reach of a stream. Is lotto a really, really random Game? An Agent Based Modelling Alessandro Perrone alex@unive.it Dept. of Economics, University of Venice       This paper describe a simulation of the Lotto Game, and it tries to give an answer to the common question. Is the lotto a random game? Lotto is game of chance such as superenalotto, powerball, and the creators of those games have gone to great lengths to make the outcomes of those games random.       If those numbers were truly random, there would simply not be a good mathematical way to gain an advantage. What I are aiming to find is an opportunity to discover flaws in the designers' schemes to make random numbers, or to discover if there are some range of values which permits to gain an advantage against the lottery.       The simulation has been written in objective-c using Swarm libraries, The Agent Based architecture is particularly suitable to write this kind of simulations, in which agents interact among them in an environment. In the simulation there are a lot of agents (from 5 to 50 different agents), each of them has it's own stategy. Every extraction has been stored in a mysql database, and they can be retrieved by the lotto agents whose rule is to get the extraction and pay the wins.       The paper starts with a short Lotto History, Chronicles Of The Game, a brief overview to most famous lottery games, then a description of some financial data about this game, and finally there's a description of the model, concentrating the efforts to comment the different strategies of the agents (there are strategies of a wide range, from the random number to play each estraction to a genetic algorithm). Can Swarm Based Systems Outperform Other Methods in Training Neural Networks? Danil V. Prokhorov dprokhor@ford.com Ford Motor Company       Particle swarm optimization (PSO) has been applied to a wide variety of problems since its inception in 1995 [1]. Yet, it seems to be a deficit of applications of PSO to neural network training problems, especially in cases of medium- and large-size networks (more than 1000 weights), large training data sets (more than 100,000 data vectors) and recurrent neural networks.       We are interested in efficient training methods for neural networks, especially those methods that scale well to problems requiring large data sets and recurrent neural networks. We have developed the training methods based on the extended Kalman filter (EKF) algorithm and applied them successfully to many problems in system modeling and control using neural networks [2]-[4]. The EKF methods operate fundamentally in the pattern-by-pattern mode of data presentation (as opposed to the PSO which operates in the batch mode), although presenting training data in mini-batches (streams) has been found to be very effective [2]. The EKF training complexity scales roughly as O(square of number of weights).       Recently, there have been claims of superior behavior of PSO applied to simple neural network training problems (see, e.g., [5]). On the contrary, our own research demonstrates that, while the PSO may be effective in comparison with simple gradient based algorithms like the standard gradient descent and other first-order techniques, it is substantially inferior to the EKF and, possibly, other more advanced methods, especially when dealing with complex problems like ones discussed in [3]. Having much more experience with the KF based techniques than with the PSO, we might well be unaware of the right set of tricks swarm researchers employ to deal with large-scale optimization problems. However, it is also possible that that, at least partially, the reason behind the observed PSO disadvantage lies least partially, the reason behind the observed PSO disadvantage lies in its batch mode of operation and poorly understood initialization of particles for large optimization problems.       We wish to discuss our comparative results with those presented in [5] and offer to future PSO benchmark studies a couple of challenging problems for training recurrent neural networks already efficiently solved by the EKF. [1] J. Kennedy, RC Eberhart, and Y. Shi. Swarm Intelligence. San Francisco, Morgan Kaufmann, 2001. [2] Feldkamp and Puskorius, "A Signal Processing Framework Based on Dynamic Neural Networks with Application to Problems in Adaptation, Filtering and Classification," Proc. IEEE, Vol. 86, No. 11, pp. 2259-2277, 1998. [3] Prokhorov, D., Feldkamp, L., and I. Tyukin, "Adaptive Behavior with Fixed Weights in Recurrent Neural Networks: An Overview," Proc. of International Joint Conference on Neural Networks (IJCNN), WCCI'02, Honolulu, Hawaii, May 2002. [4] D. Prokhorov, G. Puskorius, and L. Feldkamp, "Dynamical Neural Networks for Control," in J. Kolen and S. Kremer (Eds.) A Field Guide to Dynamic Recurrent Networks, IEEE Press, 2001. [5] Gudise, V. G. and Venayagamoorthy, G. K. "Comparison of particle swarm optimization and backpropagation as training algorithms for neural networks." Proceedings of the IEEE Swarm Intelligence Symposium 2003 (SIS 2003), Indianapolis, Indiana, USA. pp. 110-117, 2003. A Comparison of Geometric Algorithms and Agent Based Models for Fractal Simulations of African Settlement Architecture Ajith Rao mulkya@rpi.edu School of Architecture, Rensselaer Polytechnic Institute Ron Eglash Department of Science and Technology Studies, Rensselaer Polytechnic Institute       This paper describes our ongoing attempts to simulate the layouts of African settlements using a fractal approach. Many examples of African indigenous architecture have been shown to demonstrate elements of fractal design (Eglash, 1999). The settlement selected for consideration in this study is in the city of Logone-Birni in Cameroon. An examination of an aerial view of the settlement layout shows distinct fractal characteristics, such as that of self -similar scaling. These characteristics made this settlement a prime example for study of fractal based design. Among the structures, the most interesting one is the palace of the chief (Figure 1) , which shows the clearest discernible fractal appearance - e.g. the scale of the units seem to decrease as they move towards the center.       Local inhabitants reported that different reas ons were responsible for the designs implemented in this settlement. These ranged from patrilocal residency (sons would build new houses adjoining their fathers houses, hence the growth of these buildings would happen through a process of accretion) to military defense (the structure, with its series of parallel walls, would serve as an effective defense against invaders). It is interesting to see how these considerations manifested themselves in the architecture, the results of which were fractal forms. Thus, the motivation behind this study was to uncover rules that could approximate the fractal layouts in these settlements. We explored computer simulations of fractal shapes using different techniques, to examine whether they were able to reciprocate, to a degree of certainty, the layouts of these settlements.       Three distinct approaches were adopted in simulating these layouts. They were the 'transformational geometry' approach, the 'growing edge' approach, and the 'agent based' approach.       In the first approach, a seed shape was iterated to the required number in one single operation, at the same time manipulating it b y operations of scaling, rotation, etc. as required. This approach was consistent with recursive affine linear transformations [cf. Flake, 2000], where iterative copies of a shape are subjected to simultaneous mathematical transformations in one step. Although this approach gave some interesting results, it was deemed unsatisfactory in view of our goals, mainly because i t was difficult to 'tune' it to simulate an image similar to the one projected by the layout.       The second approach of using a 'growing edge' helped us overcome some of the problems faced in the first one. In this approach, each copy of a unit of iteration in the simulation would have one edge that served as a vector for the next transformation. An advantage of this approach was that the succeeding units at every step were automatically determined with relation to the previous ones. This approach was able to produce the spiraling characteristics of the layout, However, it was still insufficient in its capabilities.       The results of these two approaches indicated the need for a strategy which could better reflect the complexity of the scaling architecture. The patterns were not necessarily determined by a single rule, e.g. a spiraling could start out in a particular direction, and at some point appear to reverse direction, and so on. Also, the units were not regular throughout the layout; different sizes of units appeared in between, hence the scaling was not necessarily uniform. All this defined the need for each unit to be 'aware' of its preceding unit, and also of the overall space, before it gets formed. Another possibility was that of a probabilistic model for the layouts, which allowed for changing the rules as the simulation progressed.       This third approach, which is still under development, utilizes an agent modeling toolkit, NetL ogo to simulate the layout. This strategy involves agents which multiply on a grid according to certain rules. Using different types of agents accounts for the non-regularity of the shapes that are seen in the layout. Preliminary investigations show that fractal characteristics can emerge from this growth process. Future extensions planned for this strategy involve incorporating external constraints in the developments of these layouts. Evolving a generic model of this nature could mean that further investigations can be conducted on other indigenous architectures from different parts of the world, which is a future goal of this project. References: 1. Eglash, R. African Fractals: Modern Computing and Indigenous Design . New Brunswick: Rutgers University Press 1999. 2. Flake, GW. The Computational Beauty of Nature, MIT Press, MA 1999. Agent-based modeling vs. Equation-based modeling of Endemic Infections? Thomas Riggs triggs@beaumont.edu Center for the Study of Complex Systems, University of Michigan       Vaccine trials involving day-care centers where entire units are randomly assigned to vaccination vs. no vaccination are often used to assess transmission effects of vaccines. To estimate statistical power in such trials, one needs a prior estimate of both prevalence and the probability distribution of the number infected within the unit. Agent based modeling using Ascape was contrasted with equation based modeling using the Kolmogorov equations to estimate the prevalence of endemic infection levels in small transmission units. Using the equation based approach and assuming constant values for the (i) outside force of infection, (2) the recovery time from infection and (3) the internal contact rate, the equilibrium values for prevalence and for the probability distribution of the number of infected within the unit were calculated.       An agent-based model based on keeping the same three parameters constant and matching prevalence levels and outside force to those of the equation-based models demonstrated that the probability distributions matched very closely for the two methods. In the agent-based model, the three parameters were varied individually in a range that would cause mean prevalence to vary ( 5-10 % at the extreme parameter values. When either the outside force of infection or the contact rate was randomly varied, the mean prevalence was the same and the probability distributions were unchanged compared with a constant value for these two parameters. However, the same degree of change in the recovery time introduced a significant bias to decrease the mean prevalence. Precise estimation of the recovery time appears to be more critical than estimation of the outside force of infection or the contact rate for modeling prevalence and distribution of outcome for small transmission units. Site Evaluation and Cultural Influence Extensions to a Residential Location Model Derek T. Robinson dtrobins@umich.edu School of Natural Resources and the Environment and Center for the Study of Complex Systems, University of Michigan       Why do residential development patterns exist the way they do today? What are the factors driving and constraining residential site selection? A number of theories and mechanisms have been proposed to answer these general questions in residential location theory. The purpose of the presented research is to extend work in this field by introducing mechanisms related to heterogeneous site evaluation and cultural influence. Patterns of clustering in urban development patterns are analysed to determine if the introduction of these mechanisms cause significant differences in the spatial settlement patterns of residents, why they occur, and what these differences mean. While site evaluation and cultural influence mechanisms are topics of influence in a number of other fields, they have yet to be incorporated into residential location models. Therefore the proposed mechanisms may provide new types of explanation for why residential patterns emerge the way they do. Because the proposed mechanisms extend those that already exist they may be thought of as complementary rather than in competition with them. (poster in pdf) GIS Vector Graph Objects in Swarm Yuya Sasaki slk1r@cc.usu.edu Utah State University Paul W. Box Utah State University       In this poster presentation, we report the GIS Vector Graph Objects developed for Swarm. It is the package of Swarm object classes that employ graph algorithms to realize the importation of GIS vector data into two-dimensional raster space. This package enables the real world network data be it traffic, hydrological, or communication to be loaded in the Swarm agent-based simulation environments. While the conceptual basis is graph, the general grid acts as the platform of visual media so that it will be algorithmically compatible with Swarms cellular environment. The package consists of graph, arc, vertex, and agent classes, where agent objects can represent vehicles in traffic networks, or suspended sediments in hydrological networks. Graph class integrates and connects the other classes to enable them to represent a single consistent network. This package can read any data format from standard commercial or free programs of GIS, simply by modifying the file input portion of the package. We demonstrate the program by loading the highway data of the US west coast region in the Swarm raster, and examine the bottom-up optimality of the autonomous vehicles by embedding Q-learning modules in them. As an additional feature, each arc is assigned its free flow travel time (zero-flow cost) and marginal travel time of vehicles (marginal cost) to cause congestion as agents rush to a portion of the network. In this example application, agents learn to produce the traffic user equilibrium by bottom-up processes. Reference: http://cc.usu.edu/~slk1r/technology/vectorgis.html Spatial Evolutionary Dynamics: Spatial Replicator Equations and Cellular Autoregressive Methods Yuya Sasaki slk1r@cc.usu.edu Utah State University Paul W. Box Utah State University       We propose a version of replicator equations that incorporates the spatial interactions in cellular model. The extent of one-time-step interaction is exactly the range-one von Neumann neighborhood (NNR1). The state of a cell enters the payoff/fitness functions for all the cells in its NNR1, and this drives the spatial replicator dynamics defined either by differential or difference equations. Under this formulation, the simplex is invariant for each cell. The NE and ESS under spatial replicator dynamics remain the same as those of the standard replicator dynamics, but with far more strict conditions. While the dynamics can be specified by a system of m by n laws of motions where m is the number of spatial units or cells and n is the dimension of simplex, one will find it appropriate to design it with cellular automata. The weakest condition for the asymptotic stability of a state can be defined by the usual dynamic analyses that use eigensystems and/or Lyapunovs theorem. But this weakest condition will lack the intuitiveness for the sophisticated spatial interrelations. Thus, we adopt an alternative method to measure the evolutionary progress under the spatial dynamics. Cellular autoregressive model was modified with maximum likelihood method to let the spatial autocorrelation account for the states' deviations from ESS instead of mean. With this, the residuals represent the overall deviations from the ESS, thus enabling the whole-range evolutionary stability to be represented by zero autocorrelation and zero residual. As this result is not so frequent, we will define the state with stationary spatial autocorrelation and stationary residual as the spatially heterogeneously evolutionary stability. For this purpose, we employ the multi-layered cellular automata development environment with the aid of Swarm tool kit. Reference: http://cc.usu.edu/~slk1r/geography/spatialevolgame.html Development Report: Incorporating a Genetic Algorithm Framework into a Multi-agent Modeling System Jeffrey Fuller fullerje@student.gvsu.edu Dept. of Computer Science, Grand Valley State University Greg Wolffe Dept. of Computer Science, Grand Valley State University       The purpose of this project was to embed intelligence into varying levels of a Swarm-like simulation system. In particular, it entailed the incorporation of the JGAP genetic algorithm framework into the RePast agent-modeling tool.       In the course of integrating the framework into the tool, we identified three separate levels at which we could assimilate evolutionary computing-based intelligence. A genetic algorithm can run at the individual agent level, used to select from among possible agent behaviors. Evolutionary pressure can also manifest at the model level, using the notion of fitness to drive optimization of agents. Finally, a genetic algorithm can be employed to generate populations of simulations, used to select towards an optimal model.       Each of these implementation levels has advantages and drawbacks, and applications for which they are best suited. In this paper, we describe the details of our implementations, characterize appropriate uses, present preliminary results incorporating our system into an existing model, and indicate potential directions for future research.

Presentation Abstracts (Presenting author's name in bold) Extended Abstracts, Full Papers and Slides, when provided, are available as links off the abstract title   Exploratory design of collective behavior Eric Bonabeau eric@icosystem.com Icosystem Corporation       Agent-based modeling (ABM) enables us to reproduce emergent phenomena in collective human (and non human) systems. With a properly validated and calibrated model it therefore becomes possible to explore the range of emergent phenomena made possible by the individual-level rules of behavior and interactions between agents. While the "forward problem" of determining which emergent pattern will result from a set of individual-level rules is greatly facilitated by ABM, the inverse problem, which consists of designing the rules to create certain collective patterns, is still very difficult for a variety of reasons including: desired patterns may be difficult to formalize, the collective-level pattern landscape may be rugged, one may not know ahead of time what kinds of collective-level patterns to expect from the individual-level rules, rule space is extremely large, etc. By combining ABM with interactive evolution, a form of exploratory optimization whereby a human observer provides the (subjective!) objective function, it is possible to explore the space of emergent collective-level patterns with a view to designing "interesting" patterns. The combination of ABM with interactive evolution will be demonstrated using a simple game that can be played by a group, small or large, of human beings. Co-authors of this work are Pablo Funes and Belinda Orme, both at Icosystem Corporation. Agent Based Models of the Acute Inflammatory Response: Update on Development and Future Directions Gary An, MD Docgca@aol.com Department of Trauma, Cook County Hospital      Rationale: The Acute Inflammatory Response (AIR) is the body's first response to injury and infection. However, improvements in medical care over the past 30 years have "uncovered" a pathologic state of the AIR: Systemic Inflammatory Response Syndrome (SIRS)/Multiple Organ Failure (MOF)/Sepsis. In this situation the AIR behaves in paradoxical fashion. In many ways it acts as a complex, nonlinear system, with non-intuitive responses to manipulation. An example of this property is the difficulty translating basic science knowledge of the underlying mechanisms of the AIR into effective clinical regimes for SIRS/MOF/Sepsis; only one therapy over the last 30 years has demonstrated any statistically significant benefit. Over the past 4 years we have been using Agent Based Modeling (ABM) to try and bridge the gap between the basic science information and the clinical setting. Presented here are a series of preliminary, abstract models that demonstrate the potential benefits of this approach, as well as suggestions for future directions of investigation.       Methods: The current platform for development is StarlogoT. There are three types of ABMs presented here. The first is a global, systemic ABM that reproduces the general dynamics and behavior of the AIR. Data sources for the development of the global ABM were review articles on the components and mechanisms of the AIR. The second is a modification of the base global ABM to a specific pathogen, namely B. anthracis. The ABM simulation of Anthrax was derived from review articles on the pathophysiology of B. anthracis infection. The third is an ABM of a basic science "wet lab" model, an epithelial cell barrier function model. The basic science ABM was derived from the published papers that used the specific cell culture model simulated.       Results: The global ABM qualitatively reproduced the behavior of the AIR, as well as the unsuccessful results of the anti-cytokine trials of the 1990s. Furthermore, hypothetical regimes are shown to demonstrate the utility of ABM in designing and testing potential therapies. The Anthrax ABM demonstrates the difference between cutaneous and inhalational Anthrax as well as the effects of potential anti-exotoxin therapies. The basic science ABM of epithelial barrier function reproduced the findings of the experiments in the published papers, and may be used as an example of how potential, modular ABMs can be used in collaborative, distributed efforts of ABM development.       Conclusions: ABM is a technique of analysis that may aid the translation of basic science research into more effective clinical regimes for the treatment of SIRS/MOF/Sepsis. ABM is intended as an adjunct to more traditional research activities. ABM may be useful as a pre-testing platform for proposed clinical trials. The process of developing ABMs may lead to greater understanding of a "Theory" of SIRS/MOF/Sepsis. Finally, a "freeware" ABM of the AIR may have use as a functional, synthetic repository of basic science knowledge of the AIR. (presentation in ppt) Nobility and Stupidity: Modeling the Evolution of Class Endogamy Theodore C. Belding Ted.Belding@umich.edu Center for the Study of Complex Systems, University of Michigan       Class endogamy is a phenomenon in which nobles only marry other nobles and commoners only marry other commoners. The origin of class endogamy, and of social stratification in general, is a major open question in archaeology. This paper implements a verbal model proposed by Marcus and Flannery as a class of agent-based computer models by generalizing and simplifying a mathematical model of marriage markets developed by Burdett and Coles. One force that can produce class endogamy occurs if agents are only willing to marry suitors having status no less than some fixed value below the status of their highest-status suitor, which they can learn. Another such force results if children inherit the average of their parents' statuses. In contrast, status achieved over an agent's lifetime can be viewed as noise, analogous to mutation in biological evolution. I propose that class endogamy may have resulted from forces such as these, along with other factors such as ideology. Simulation results are presented, and potential areas for future research are sketched out. The validity of these models for any particular culture depends, of course, on whether these forces were actually operating in that society. (paper in pdf) Inferring Individual Behavior with GP: Agent-Based Models of Public Goods Provision. Riccardo Boero R.Boero@surrey.ac.uk Dept. of Sociology, University of Surrey       Agent Based Models are useful for the process of understanding dynamics and causal relations of complex social systems. That kind of modeling approach is based on the process of reducing real complexity to the model micro specification and of studying its outcomes via computer simulation. But to increase the scientific value of the research process, a deeper empirical foundation of the micro specification must be developed. In the paper, I report some results obtained searching for nano foundations of socio-economic ABMs, i.e. empirical foundations of the micro specification. In particular, an attempt to use data collected in classroom experiments is presented and a methodological procedure is evaluated. In fact, the paper focuses on an example based on some experiments about public goods provision, showing how some parts of the micro specification can be easily founded on reality (i.e. how the interaction and endowments structure can be made explicit as experimental ones) while individual behavior is problematic. In fact, a complicate procedure is needed to infer a behavioral strategy useful for modeling and scientific purposes. Thus, an inferring procedure using Genetic Programming is presented and, finally, some models are presented too with the aim of stressing how the proposed procedure helps building models to understand such kind of social and economic dilemmas. SumWEB: stock market experiment environment for natural and artificial agents Alessandro Cappellini cappellini@econ.unito.it University of Turin, Italy       SumWEB (SUM Web Economic Behaviour) is a Swarm based web-application.       The simulation core is based on a Objective-C stock market simulation, SUM (Surprising (Un)realistic Market), developed by Pietro Terna (Terna 2000) enriched by new features.       The prices formation and the order enqueuing rules were directly inspired by MTA (Mercato Telematico Azionario, the Italian Stock Exchange). We included rules such as opening and closing auction, but more relevant a tick by tick formation price mechanism.       In order to increase simulation realism we can create various books to add more than one equity stock. The stock prices are used to calculate an index so we can activate a special book to collect index future proposals.       The agent population is very rich and heterogeneous. We can activate minded agent (equipped with an artificial neural network) or no minded purely random agent. We designed also some agents based on simple trading rules such as the stop loss agent or the arbitrageur agent.       SumWEB was mainly designed to introduce humans inside the simulation using the avatar tecnique. With this simple idea we can build a bridge from pure ACE (agent-based computational economics) approach to experimental economics. Humans can stress test model, can reveal implementation or logical error, slackness, or better can show unpredictable behaviours. We can study humans behaviour in an artificial controlled lab.       SumWEB was electively used for two experiment, organized by Faculty of Economics (University of Turin). The first one was a class game with 57 persons on 6th May 2003 for one hour. The second one was a two weeks long online experiment, from 8th to 21th May 2003, with 152 persons over 486 agents. In both experiment the market was populated by three equity stocks and one index future. slides in pdf Swarm and GNUstep Development Report Scott Christley schristl@nd.edu University of Notre Dame       In this report, I discuss the development efforts to allow Swarm applications to work within the GNUstep environment; the result is that Swarm acts like a third-party library within the GNUstep environment allowing Swarm to utilize existing GNUstep functionality.       GNUstep's website at http://www.gnustep.org describes GNUstep as 'a free, standard, object-oriented, cross-platform development meant to provide generalized visual interface design, a cohesive user interface, and look good as well. GNUstep is based on and completely compatible with the OpenStep specification developed by NeXT (now Apple Computer Inc.). GNUstep also implements many additional classes and methods, some from the Cocoa API for the sake of compatibility. GNUstep is written in the object-oriented language "Objective-C", a superset of C which adds object-orientation to C.' The development project has several milestones and long-term goals: Remove usage of Swarm's internal libobjc and use the standard Objective-C runtime provided with the GNU compiler. Package Swarm as a third-party library within the GNUstep Makefile system so that it can be asily incorporated into GNUstep applications. Have Swarm's simulation scheduling mechanism work alongside GNUstep's graphical event processing. Replace Tcl/Tk/BLT graphical interface with a GNUstep one. Replace some Swarm internal implementations with GNUstep classes. Allow Swarm classes, both internal and user-defined agents, to be a palette in GORM. Creation of a Swarm application type within ProjectCenter to provide a set of template files with some standard classes, menus, and windows pre-defined for the user. The current progress of these milestones, any issues encountered during the development process, work still to be done, and some exciting future prospects will be presented. slides in pdf Agent-Based Modeling and Simulation of Strategic Scenarios with Repast 2.0 Douglas Druckenmiller1,2 William Acar2 Marvin Troutt2 1ddrucken@bsa3.kent.edu 2Graduate School of Management, Kent State University.       This paper provides a background discussion of agent -based modeling (ABM) and simulation of strategic scenarios. Causal mapping is introduced as a structured method for situational formulation and analysis of unstructured strategic problems. Causal mapping includes specific processes and analytical approaches offering cognitive modeling support for problem formulation and scenario planning. A prototype system for the development and simulation of causal maps using RePast 2.0 is described. A typical application is described and implementation issues are discussed. The prototype system provides the development of a human-artificial conceptual map for assumptional analysis of strategic scenarios that serve as the basis of selecting relevant information for strategic decision making. (paper in pdf, poster in ppt) A Spatially Explicit, Bioenergetically Cosntrained, IBM of Predator-Prey Interactions in a Stream Eliezer Gurarie eliezg@u.washington.edu Quantitative Ecology and Resource Management, University of Washington       The northern pikeminnow (Ptychocheilus oregonensis) is a fresh-water predator that contributes significantly to the mortality of ocean-bound juvenile salmonids (Oncorhynchus sp.) in the lower Columbia River basin. Models used to estimate predation on salmonids tend to lack information about spatial variability of predators and prey and complexity of the environment and to ignore the energetic constraints of feeding fish.       A spatially explicit bioenergetically constrained individual based model of pikeminnow predation on salmon smolt was developed in SWARM. In it, a predator with basic foraging behavior encounters passing prey. The model provides the opportunity to explore environmental variables (temperature, flow velocities, light availability), and individual behavioral variables (reaction distances, aggregation of prey) in a unified context. Simulations show that growth and consumption display a strong though qualified dependence on temperature, spatial structure of migrating prey, and prey density.       The model has potential for testing the assumptions used in smolt migration and survival models and exploring the role of heterogeneity and environmental complexity on the pikeminnow-salmonid system. An expansion of this approach would integrate aspects of visual foraging, bioenergetics, swimming mechanics, behavioral responses and hydrodynamics, and contribute to a unified theory of predator-prey interactions in aquatic environments. The Immune System as a Complex Adaptive System: A RePast Simulation of the Anti-Viral Immune Response. Virginia A. Folcik, Ph.D.1 vnivar@hotmail.com Charles G. Orosz, Ph.D.1,2 orosz-1@medctr.osu.edu 1Department of Surgery/Transplant, The Ohio State University College of Medicine and Public Health 2Corresponding Author.       The immune system is a prime example of a complex adaptive system, with Individual cells that follow rules for behavior based upon detection of signals and contacts with other cells in the environment. We have created a simulation of a human anti-viral immune response using the RePast software framework. The agent-based simulation includes three windows that represent a generic tissue site with parenchyma that becomes infected with virus, a lymph node site with cells that can become activated to fight the viral infection, and the peripheral blood that carries the responding immune cells and antibodies back to the site of infection. The simulation uses seven agent types and twenty signals to represent Parenchymal Cells, B-Cells, T-Cells, Macrophages, Dendritic Cells, Natural Killer Cells and the virus, and pro- and anti-inflammatory cytokines, chemokines and antibodies that such cells use to communicate with each other. The numbers of agents present as well as the quantity and types of signals present depend upon rules for proliferation and the release of cytokines that the agent types follow. Individual agents have various states, migrate from one window to another and live or die as the rules for their behavior dictate. A typical run of the simulation involves the entry of initial conditions (ratios of immune cell types), then the execution of the simulation during which the numbers of agents and quantities of signals are recorded. Given sufficient time, the outcome of a run may be either that the virus infects all of the parenchymal cells resulting in the death of the tissue (a viral "win") or the elimination of the virus and all virally infected cells with regeneration of healthy cells and restoration of the tissue to equilibrium conditions (an immune system "win"). Consistent with the theoretical properties of a complex system, our experiments have found initial conditions that always produce the same win/loss results, but the profiles of cell proliferation and signal production that occur are unique for every run of the simulation. Other initial conditions have been found that produce varying win/loss ratios.       We plan to be able to use our simulation to explore formative patterns of agent behavior that develop within a complex adaptive system, to evaluate how information is used for decision making as responses evolve, and to develop methods of generating and evaluating simulator data that can be used to identify the strengths and weaknesses of clinical and experimental tools that are currently in use. slides in powerpoint Coupled Eulerian-Lagrangian Agent Individual-based Modeling (CEL Agent IBMs) for Fish R. Andrew Goodwin rag12@cornell.edu Environmental Laboratory, US Army Engineer Research & Development Center at Portland District James J. Anderson School of Aquatic & Fishery Sciences, University of Washington John M. Nestler Environmental Modeling & System-wide Assessment Center, US Army Engineer Research & Development Center Larry J. Weber IIHR Hydroscience & Engineering, University of Iowa       We describe a theoretically- and computationally-robust mathematical method for decoding movement patterns of individual fish responding to biotic and abiotic stimuli in 3-D space-time. The method, coupled Eulerian-Lagrangian agent individual-based modeling (CEL Agent IBM), integrates the three primary theoretical frameworks for mathematically describing the movement of animals: Lagrangian, Eulerian, and agents. CEL Agent IBMs integrate a Lagrangian particle-tracking algorithm supplemented with behavioral rules into a 3-D Eulerian computational fluid dynamics (CFD) model. Behavioral rules derived from an agent-based, event-driven foraging model query stimuli information from the CFD model or a priori field data collected in a dense grid. Back-casting simulation analysis results in a mechanistic, mathematical formulation amenable to accurate forecasting. We demonstrate the utility of the method by presenting results from a CEL Agent IBM application at Lower Granite Dam on the Snake River, Washington, USA in which the observed 3-D movement and passage patterns of downstream migrating juvenile salmon were successfully decoded with sufficient accuracy to assist engineering design. The prototype CEL Agent IBM, the Numerical Fish Surrogate, explains 74% (r2 = 0.74) of variation in fish passage for eleven different structural and operational configurations that vary substantially in flow and bypass system design. For comparison, colored dye (or passive particles) released from the same locations yields an r2 of 0.53. The Numerical Fish Surrogate is used by the US Army Corps of Engineers to study observed fish movement, evaluate behavior hypotheses, and forecast plausible fish movement and passage response to alternative designs of bypass structures at federal hydropower dams. CEL Agent IBMs can be generalized to other aquatic or terrestrial ecosystems in which the behavior and movement of individuals is important. Discrete Evaluation and the Particle Swarm Algorithm Tim Hendtlass Centre for Intelligent Systems and Complex Processes, Swinburne University of Technology Tom Rodgers Centre for Intelligent Systems and Complex Processes, Swinburne University of Technology {thendtlass,trodgers@swin.edu.au}       We propose that the optimal performance of the PSO algorithm should differ from that of the real life creatures on which PSO is modelled. If a bird finds a good food source, the likely behaviour for a flock is to congregate there, settle and feed. However, once PSO has found an optimum, while some particles should explore in the immediate vicinity for any better optimum present, the rest of the swarm should set out to explore new areas.       The common PSO practice of only evaluating each particle's performance at discrete intervals can, at small computational cost, be used to automatically adjust the PSO behaviour in situations where the swarm is 'settling' so as to encourage part of the swarm to explore further. paper in pdf Agent Based Models of Competitive Sympatric Speciation: An Investigation into the Role of Mate Search Tactics and Complex Phenotypes Rainer Hilscher rainer.hilscher@gmx.net Affiliation Evolutionary and Adaptive Systems Group, Informatics, University of Sussex, UK       Competition can be a creative force. The evolution of new species due to competition for a limited resource between individuals of a population is one example that has received increased theoretical attention in evolutionary biology. Competition is one form of disruptive selection in sympatric speciation models. In contrast to allopatric speciation where geographical isolation enforces reproductive isolation between new newly formed populations, there is no barrier to gene flow in sympatric speciation. Competition and the subsequent evolution of assortative mating can generate such a barrier.       After a first phase of studies that showed that sympatric speciation does work in principal attention is now slowly shifting towards understanding the nuts and bolts of this species creation process. Published analytical and simple individual based models, however, are very limited in biological realism due to their complexity constraints. Individuals are not treated as single units, resources (such as food) are not explicitly modeled and many other simulation features are also highly simplified. Mate search tactics in the context of assortative mating and complex phenotypes are just two domains that fell prey to mathematical intractability in analytical models and simple IBM's. Here I present an agent based model of competitive sympatric speciation that does not suffer the aforementioned complexity constraints. One reason lies in the very nature of agent based modeling and the other resides in the implemented architecture of the model. By following a plug-and-simulate approach (similar to what Gulyas calls Relational Agent Models) based on interfaces and dynamic class loading it is possible to test many different scenarios by simply exchanging a few simulation objects. Several different, spatial and non-spatial ecological conditions (food distributions, e.g. 2-patches, gaussian and gradient) can be tested against different implementations of mate search tactics (best-of-n and threshold-based variants). Results indicate that best-of-n mate search is significantly more successful in splitting a population than threshold-based search.       A second line of investigation refers to complex phenotypes. In traditional models only a single fitness trait is considered. This model implements phenotypes that consist of pleiotropically linked sub-units. Agents with complex phenotypes have to compete for multidimensional food items. Results show that there exists an optimal degree of interaction between sub-units for speciation to occur. paper in pdf Agent-Based Modeling of Cultural Change in Swarm Using Cultural Algorithms Ziad Kobti kobti@uwindsor.ca School of Computer Science, University of Windsor Robert G. Reynolds Department of Computer Science, Wayne State University Tim Kohler Department of Anthropology, Washington State University       The multi-agent Village simulation was initially developed to examine the settlement and farming practices of prehispanic Pueblo Indians of the Central Mesa Verde region of Southwest Colorado [1,2]. The original model of Kohler was used to examine whether drought alone was responsible for the departure of the prehispanic Puebloan people from the Four Corners region after 700 years of occupation. The results suggested that other factors besides precipitation were important. We then proceeded to add economic factors into the simulation, first allowing agents to engage in reciprocal exchanges between kin. This resulted in larger populations, more complex social networks, and more resilient systems. However, the exchange was done randomly and individuals did not remember the transactions. In this paper we explicitly embed the reciprocal exchange process within a Cultural Algorithm, where individual agents can remember individuals that they have cooperated with. Also, in the cultural space the group can learn generalizations about what kind of relative is likely to successfully respond to a request. These generalizations are used to drive changes in requestor behavior. The results of this approach produced an even larger and more complex system exhibiting greater dependence on hub nodes that are sensitive to precipitation. paper in pdf Exploring human-environment complexity: an agent model for across-scale and interdisciplinary integration. Li An lian@umich.edu Department of Fisheries and Wildlife, Michigan State University Center for the Study of Complex Systems, University of Michigan Marc Linderman Department of Fisheries and Wildlife, Michigan State University Jianguo Liu Department of Fisheries and Wildlife, Michigan State University       Traditional top-down approaches (e.g., state variable approach) to studying wildlife habitat often ignore individual-level information about the human population of interest, especially at household and/or individual level, and often cannot capture or explain some key processes. This study reports on an agent-based spatial model that addresses this issue. The rapidly growing rural population in the Wolong Nature Reserve for giant pandas (China) follows a traditional rural lifestyle, in which fuelwood consumption has been the main driver for panda habitat degradation. Following the life history of individual persons and households, this model equips the individual and household agents with knowledge about themselves, other agents, and the environment (topography, forests, etc), and allows them to interact with each other and the environment based on a set of rules obtained from our fieldwork. The agents and forests change and talk to each other over time and space, resulting in emergent human and habitat dynamics. Aside from providing insights to panda habitat conservation, this model may provide wildlife researchers with a useful tool to study how habitat patterns change over time and space as the local people, households, and forests evolve and interact with each other. The Computer Experiment in Computational Social Science Greg Madey gmadey@nd.edu Computer Science & Engineering, University of Notre Dame       The year 2003 was the 50th anniversary of the invention of the "computer experiment" by Fermi, Pasta and Ulam. The computer experiment was offered as the third way of doing science at the time. In Kuhn's normal science, the scientific method suggests the generation of new knowledge by making observations of a phenomenon, identifying curious aspects of the phenomenon, generating a falsifiable hypothesis to explain the phenomenon, and designing an expermiment to disprove the hyposthesis (Popper 1982). Should the experiment fail (to disprove the hypothesis) it is accepted as an explanatory model until eventually replaced by something better. Fermi et al proposed the use of the computer experiment for inquiry into the physical sciences where the phenomenon cannot or is not easily observed. Over the last decade various social science disciplines, including political science, anthropology, sociology, and organizational science began to embrace simulation as one method of inquiry in what is sometimes called computation social science. Recently, Axelrod (1997), McKelvey (1999), Goldspink (2002), Kluver et al (2003) and many others have explored the role of computer simulation as a source of new knowledge in the social sciences. We integrate their analysis and present another view of computer simulation as part of the classical scientific method applied to the investigation of social systems. The hypothesis of the classical scienfic method becomes the conceptual model of the social scientists, which in turn is implemented in a computer simulation. Computer experiments are conducted using those computer simulations. (slides in pdf) StarLogo Model of Crayfish Microhabitat Selection Stephen D. Morse Morses4@aol.com Biology Dept, Camden High School, Camden, NJ       Teaching how organisms interact with habitat is different from researching the problem. Research often paints a complex picture. Multivariate statistics are commonly used to describe the influence of several habitat variables on distribution. At the high school level, textbooks often paint a simpler, sometimes oversimplified, picture. I attempted a third approach by developing a Starlogo model. It uses actual field data on crayfish distribution and several microhabitat variables to illustrate habitat use in an intuitive, visual, and accessible manner. The model uses a population of artificial crayfish, reacting to habitat variables in parallel, as instructed by the student. The goal is to develop habitat selection criteria for these artificial crayfish that will place them in habitat on the computer screen corresponding to real habitat used by real crayfish at the study site. Students determine whether individual variables affect young-of-the-year crayfish numbers at the study site positively or negatively. Then they estimate the degree of crayfish response to each variable, and combine the variables to produce their best multivariate model. (slides in ppt) Diffusion of Innovations in Small Worlds: Taking Shortcuts While Seeding? Kerimcan Ozcan kozcan@umich.edu School of Business, Center for the Study of Complex Systems, University of Michigan Venkat Ramaswamy University of Michigan       Whereas most research on diffusion of innovations (Bass 1969), network externalities (Katz and Shapiro 1986), information cascades (Bikhchandani, Hirshleifer, and Welch 1992), and fashions (Miller, McIntyre, and Mantrala 1993) require and recognize network phenomena without explicitly modeling them or doing so under very restrictive assumptions, most research in social network analysis facilitates the description of the "ties that bind actors in a network" (see Wasserman and Faust 1994 for a general review) without dynamically linking structure with concrete social processes and individual manipulations (White, Boorman, and Breiger 1976). Subsequently, one has to study how information flows and other transactions relate to structural patterns and their change. This is the main objective of this paper. In particular, we superimpose a theoretical model of innovation adoption and word-of-mouth interaction at the micro-model onto the small-world insight that a very small number of random, global links at the macro-level can shrink the network drastically (Watts 1999).       We first propose a theoretical model of word-of-mouth interaction utilizing information- and decision-theoretic conventions (Chatterjee and Eliashberg 1990; Feder and O'Mara 1982; Jensen 1982; Roberts and Urban 1988). Next, using the small-worlds formalism proposed by Watts and Strogatz (1998), we generate networks that are connected, have minimal structure, are made up of unidirectional and non-valued links, and range between a topological ring and a complete graph. Then, we utilize the SWARM agent-based modeling platform to investigate the effects of network size, number and distance of shortcuts, and number and collocation of seed agents on the aggregate dynamics of innovation adoptions as mediated through word-of-mouth traffic. We close by discussing these results and what they imply for managerial practice. IDEAS - Interactive Development Environment for Agent-based Simulation Alessandro Perrone alex@unive.it Dept. of Economics, University of Venice Andrea Pellizzon University of Mathematics, University of Padua Licia Salce University of Mathematics, University of Padua       This document describes the new features implemented in the IDEAS, and interactive IDE for Agent based simulations. With this tools, building a simulation is then done by adding components from the component palette to the property pane, customizing these components by editing their properties, compiling the project and then running the resulting simulation.       This paper gives an overview of the environments showing how is simple to write simulation using different Agent based environments even if there's a little knowledge of the agent based paradygm. What's an IDE? Upon a standard definition IDE, integrated development environment is a system for supporting the process of writing software. Such a system may include a syntax-directed editor, graphical tools for program entry, and integrated support for compiling and running the program and relating compilation errors back to the source. Such systems are typically both interactive and integrated, hence the ambiguous acronym. They are interactive in that the developer can view and alter the execution of the program at the level of statements and variables. They are integrated in that, partly to support the above interaction, the source code editor and the execution environment are tightly coupled, e.g. allowing the developer to see which line of source code is about to be executed and the current values of any variables it refers to. I have, over the course of the last few years, tried just about every Interactive Development Environment out there to build the simulation. In my opinion they all share two things.       1. They try to do too much, which makes them all large, slow and painfully hard to use.       2. They force me to change the way I work. I want a development system that works the way I do.       That's why we have developed this IDE. Its development paradigm matches the standard simulators, not the other way around. Relationships between Agent-Based Models and Geographic Information Systems: An Illustrated Catalog William Rand1 wrand@umich.edu Daniel G. Brown2,1 Michael North3 Rick Riolo1 Derek T. Robinson2,1 1Center for the Study of Complex Systems, University of Michigan 2School of Natural Resources and the Environment, University of Michigan 3Center for Complex Adaptive Agent Systems Simulation, Argonne National Laboratory       Spatial data models in geographic information systems (GIS) are used to structure the (mostly static) geographic world so that it can be represented within a database. Two conceptual data models dominate GIS representations of the world, i.e., the field and object views. Spatial process models are similarly structured representations of dynamics within the geographic world. Two dominant conceptual views of spatial processes, borrowed from the Eulerian and Lagrangian views of fluid dynamics, yield models of change and models of movement. In this talk we argue that spatial extensions of object-based process models require that these process models be closely coupled with data models that can be used to explore, explain, and interpolate the spatial data that results from the process model. We discuss how alternative spatial data models constrain or enable close coupling with alternative types of spatial process models. We briefly examine past attempts to integrate spatial data and process models. We then describe how independent developments toward the object-oriented computational paradigm within both geographic data modeling and spatial process modeling provide a new opportunity for close coupling. We discuss the scientific and practical advantages of developing systems that closely couple spatial data models in the form of GIS databases with spatial process models in the form of agent-based models (ABM). The rest of this talk focuses on developing a catalog of relationships between geographic data (fields and objects) and agent-based process models, based on whether the agents have an identity association with spatial feature(s), whether or not such spatial features can move, whether or not they can change, whether or not the agents can change non-agent spatial features or be changed by these features, and whether time is treated as time steps or discrete events. These types of relationships are then illustrated with examples from our work in coupling ABM and GIS. Moreover, there is the question of implementation. We discuss several of the issues that must be addressed when actually implementing the connection between ABM and GIS. We illustrate these questions with examples from our own work and discuss our plans for further integration in the future. presentation slides in pdf Business Application of Agent-Based Simulation: Complex and Dynamic Interactions of Motion Picture Market Seung-Kyu Rhee skrhee@kgsm.kaist.ac.kr Graduate School of Management, Korea Advanced Institute of Science and Technology (KAIST) Wonhee Lee Graduate School of Management, Korea Advanced Institute of Science and Technology (KAIST)       Movie is naturally a new product and has short life-cycle from one week to several months. With huge initial investment and high uncertainty of the market performance, all constituents of the movie supply chain, from a writer with an idea to theater managers with screens to allocate, face very difficult decision problems. Given a movie to sell, a distributor has to decide how much marketing budget to spend, when to release it, how many screens to secure. These decisions should be based on the projected market performance, which, in turn, would be influenced by the decisions themselves and many other uncontrollable factors, notably the early performance of the movie itself.       Existing literature ranges from simple statistical forecasting models to a complex dynamic Markov chain model with behavioral parameter estimation. Some agent-based models have been proposed to describe the near-chaotic market behavior in terms of market share change. To our knowledge, no existing model is comprehensive enough to be useful for decision makers in motion picture industry. In agent-based simulation community, there is a tendency to prefer simple models. From practitioners' viewpoint, however, it does not help much to confirm the fact that the market is too complex and anything is possible. In this paper we expand the scope of the movie market model by including diverse sources of movie quality information and competition effect.       A movie is a cultural product, the quality of which can only be determined by experiencing, and therefore subjective. When a moviegoer has to decide whether she goes to a particular movie or not, however, she needs at least some information on the movie quality. The consumers receive information of movie quality and attractiveness from diverse sources: expert critique reviews; suppliers' marketing signals including theater trailers, previews, and advertisements; initial box office performances; and words of mouth (WOM) from friends. The quality information distribution is not uniform across the market. Some can be considered as universal and some can be partially available, and WOM is local. Also the contents can be contradictory. A moviegoer exposed to the information and with her own preference and constraints, has to decide what to do. She cannot go to all the attractive movies in a limited time period, where comes in the competition effects. We model the complex consumer decision behavior using agent-based simulation.       The simulation experiments were carried out in two ways. First we developed a baseline model and tried several scenarios to examine the propositions suggested in existing researches of motion picture industry. Second, we estimated the model parameters from actual movie data, and then we compared the actual and projected market performances. Several interesting results followed. We discuss the impacts of inherent movie quality, WOM, critique reviews, and marketing efforts to movie performances contingent upon the market competition. Next we discuss the strategic implications for movie distributors regarding to marketing intensity and release timing. Finally, the empirical validation using Korean market data show that the model generates quite close projections for both opening market shares and final market shares for four different data sets.       The model discussed in this paper is one focusing on the complex consumer dynamics. When applying agent-based simulation to a real and complex decision situation, it is more important that every additional variable and agent should be justified by increased insights and relevance. We discuss the issues by comparing model results and existing literature. Discussion on the model extension to add the theater objects and overlapping release strategies will close the paper. slides in powerpoint On the Virtues of the "Shame Lane" Matteo Richiardi m.richiardi@labor-torino.it LABORatorio Revelli, Centre for Employment Studies       In July 2003 a new Road Code was approved by the Italian parliament. Among many reforms whose validity is not questioned here, the new law states that on three-lane motorways the right lane should not be reserved anymore to slow vehicles alone. As in two-lane roads, all vehicles must now drive on the right lane, as long as it is not occupied by other vehicles. The model developed in this paper casts doubts on the validity of such a change, suggesting that reserving a separate lane for slow vehicles is generally better, in terms of number of accidents and slow-downs, than the new one. This conclusion has a general validity beyond the Italian case. Moreover, it is shown to be extremely robust to refinements of the main assumptions concerning driving attitudes and the stochastic arrival of accidents. paper in pdf Toward Constraint-Sensitive Agents David L. Sallach sallach@uchicago.edu Argonne National Laboratory       In recent years, significant progress has been made in social research based upon agent simulation methods. A baseline domain space in which agents with directlydefined deterministic and probabilistic capabilities has been extensively explored, and computational tools to support such modeling and design have been developed. Simulation based on this approach has been able to generate intuitive aggregate-level outcomes in a variety of research domains and, in various cases, have created higherlevel patterns of emergence. Thus, the initial insights and potential of social simulation have created a rising research paradigm.       Notwithstanding early achievements, the new paradigm has much unrealized potential as well. Social processes are subtle and complex, observably fluid and volatile. Algorithms, exogenous rules and mean field approaches are unable to capture their creativity, innovation and unpredictability. The use of tag-based functions to model cultural processes is a suggestive example of the limitations of discrete models of social dynamics.       To realize the potential of agent models in social research, it is important to construct agent capabilities based on higher-order computational processes. There are various ways in which this might be incorporated into model design. First, chaotic models might be incorporated as a direct aspect of agent capabilities. Second, some combination of deterministic, probabilistic and/or chaotic computation can be used to construct second or higher-order possibilities. One higher-order computational model that can be used is constraint modeling. The prospective implications of using constraint modeling to build agent models is considered in the present paper.       The use of constraint programming originated as a way of solving optimization problems. Many social patterns originate as actors intending, acting or interacting under constraint. Examples include ecology, social structure, mutually constituted intelligibility, maintenance of the social self, reflexive accountability and/or multiple intentionality states and, of course, these are not exhaustive. Underlying such constraint mechanisms is the possibility that they are endogenous as well as exogenous, i.e., that relevant agents are not only shaped by exogenous constraints but also, by their communications and actions, they shape the constraints under which subsequent events occur.       Toolkits designed to support constraint models will require special components and features. The presentation will conclude by identifying prospective tool desiderata. Psycho-Computational Models of Human Linguistic Development: Applications for Swarm Teresa Satterfield tsatter@umich.edu Dept. of Romance Languages and Literatures and Center for the Study of Complex Systems, University of Michigan       In this updated version, SWARM modeling is used to examine premises of linguistic theory and first- and second-language development. Hypotheses have been offered for how languages in historical language contact settings acquired their native-speaking communities. A long-standing challenge for these studies is how to reliably test the claims put forth, since it is difficult to provide a complete account of the origins of such conta