Livrables
We have built a simulation environment for large-scale agent-based simulation. First of all, the simulation environment facilitates implementing simulation models by means of reuse of software components. Second, it allows running a series of simulation experiments with different models, different model parameters and different implementations of particular components using a specification mechanism (SM). Third, large-scale simulation is supported through parallel execution on clusters of workstations. For each new model, there are two components that need to be implemented. The first is called agent collections and landscape elements, which scales to hundreds of thousands of agents. Collections of agents and landscape elements can be distributed over a set of workstations, giving way for large-scale parallel simulation.
Designing a software component for these distributed collections is supported by a library of convenient primitives. The second is called worker, which is configured from a set of behavioural components to implement the behaviour of agent collections. During parallel simulation, there are several workers each simulating its partition of agents. These two components for the time being need programming skill. One can have different implementations of those components for the same model. The SM for component dependencies, output channels, and simulation input parameters are used to configure the experiment. The structures of the SM are simple to write. The problem is the analysis and visualization of the output data. The most useful approach is to use one of the available commercial packages, e.g., MATLAB. The simulation environment has also a provision for specifying the models by means of a behaviour language for agent-based simulation. A model specification, consisting of descriptions of agents and their behaviours, and of the landscape where agents reside, could be compiled into the data collection modules and worker modules, using libraries provided by the simulation environment.
Information cities project defining issue was the need to respond to the basic question of how the web economy organizes its use of (virtual) space: explain evident concentrations of population and economic activity over the web (information cities) and illuminate fast growth rates obtained by leading web locations. In this project we have modelled the self-reinforcing forces that drive growth and prosperity in the web ecosystem and support the process of formation and evolution of information cities. We show that the strong organization of the web economy around a handful of locations is clearly something that has emerged, not necessarily because of any inherent initial qualities of different web sites (is http://www.amazon.com so different from http://www.barnesandnobles.com) but rather through self-organization involving various feedback mechanisms.
In fact, the picture that results from our theories and models, based on both, i) appreciative (empirically observing key-Internet players' behaviours) and, ii) formal theorizing (modelling these behaviours within a "built-from-scratch" computational web economy and running multiple experiments) demonstrates that:
All these concentrations of population in particular web locations form and survive because of some form of agglomeration economies (fuelled, in most cases, by informational increasing returns), in which concentration and growth themselves create the favourable environment that supports further or continued concentration and even faster growth. The above conclusion was reached by developing a modelling framework, which can account for both strong patterns of growth obtained in the web economy (concentrations of population and economic activity over the web, and, fast growth rates obtained by leading web locations). Research and development of such a modelling framework was the primary strategic objective of iCities Project.
The developed modelling framework i) follows a "bottom-up" approach: web sites' growth and population agglomeration, emerge endogenously from the behaviour of individual agents (i.e. web sites and Internet visitors) and the various interactions between them, and ii) takes into account the agglomeration economies present in the World Wide Web. In this framework we have developed three particular agent-based models and a simulation infrastructure to satisfy the particular requirements of our models.
First model's starting question is simple: Why cities, information cities, i.e. large agglomerations of people and economic activity emerge in the virtual world, and why web economy exemplifies particularly fast growth rates for its elements. The model provides a sound basis for the dynamics of population concentration in the Web and put forward an explanation to web sites' exceptionally high growth rates, by developing a simulated web economy, with behavioural and economic variables. The model i) reproduces the empirically observed power law distribution of Internet users across web sites, and ii) suggests that the existence of particularly strong, and rapidly formed, agglomeration patterns in the web involves a specific growth process that emerges from various forms of informational increasing returns.
The second model has the main objective to examine evolutionary dynamics in the formation of the web geography and economy by studying an experimental very complex environment. From various model simulations, we have been able to investigate rules of competition and success strategies within this experimental web economy. More precisely, we have obtained fast growth pattern for web leaders (largely depending on the structure of the networks carrying agents' decisions) and co-existence of highly diversified, partially diversified and specialized sites.
The main objective of the third model is to frame the forces that drive the phenomenon of aggregation of merchant web sites (B-to-C) competing in a differentiated electronic market where search cost (costs related to find a good in a differentiated market) for the consumers are independent of their adaptation/transportation cost (when the good they find does not match with their preference).
The simulation infrastructure (large-scale agent-based simulation environment) has been built to provide opportunities for experimentation and progressive formation of global dynamic behaviours. The environment has two versions, a sequential executing on a single workstation, and a parallel executing on a cluster of workstations. Multiple experiments showed scalability and speedup when executed in both clusters of workstations and multiprocessors machine.
In iCities project we develop and study the behavior of non-analytical models with emerging and self-organizing properties. Our models simulate a web economy where patterns of organization (mostly, aggregation of visitors within particular web sites behaving as information cities) emerge within two populations of boundedly rational agents, Internet consumers and web sites, who interact locally or globally. In particular, we study problems of "product-choice" (i.e. web-site choice) in an environment characterized by increasing returns and various other behavioral rules. Placing agents with particular behaviors in such an environment proves sufficient for the emergence of various aggregation and segregation structures. Such structures have characteristics that are suggestive of information cities. Agent-based computational models are used to study the non-equilibrium dynamics, in which aggregated structures are perpetually born, growing, prospering or diminishing.
Our approach consists to observe model's outcomes and conclude on their properties, stability and their dependence on initial conditions and other key-parameters, to "measure" aggregation trends, to evaluate how this "aggregation index" varies as a function of different model parameters and compare model's output to aggregate statistical data. The Tool provides all the necessary functionality to monitor the above-described tasks, to visualize interactions among models' populations and highlight emerging structures.
ICities project has the objective to run various simulation experiments (by modifying key model parameters or by playing scenarios using various combinations of simulation infrastructure's models) and observe emergence and stability of aggregated structures. Essentially, the Visualization Tool helps the Model Analyst to proceed with information obtained from the experiments. It gets basic functionality from MATLAB to elaborate and accordingly present data, which are imported, after each experiment, from the iCities simulation platform built on top of Mozart/Oz.
Each simulation experiment generates a number of output files, which include statistical information, for example the number of users that visit web sites at each time step. The Model Analyst generates in each experiment several output files according to experiment's needs and the specific model under study. Files that are generated from the simulation environment and saved from the Model Analyst according to experiments' requirements are read inside MATLAB. A mat-File is generated for each experiment. The Model Analyst can then launch the main figure of the visualization tool, load the mat-File that relates to the experiment, and start the analysis. The functionality and the user interface of the visualization tool main figure changes as a function of the model under experimentation.
The components (pop-up menus, pushbuttons, edit-boxes) of the Visualization Tool User Interface relate to information data for the experiment the Model Analyst is studying. The pop-up menus carry information for the name and the version of the experiment. Edit boxes display experiment's parameter values and pushbuttons compute and plot statistical information for the loaded experiment. Model analyst can easily load a second experiment and load it in the Visualization Tool then, both experiment files are available to the Model Analyst for comparative analysis. The User Interface and the functionally it provides, automatically change depending on the loaded experiment version. Simultaneous analysis of different experiments can be easily performed by using the pop-up menu that displays experiment's name and selecting, from a list of names, one particular experiment among the loaded experiments.
By clicking to a pushbutton, the Model Analyst obtains a figure that displays a particular statistical graph for the selected experiment. For example a figure that displays the histogram of the occurrences of sites by popularity, etc.
Extensions of the Visualization Tool can be used for the analysis of any discrete time agent based computational model as i) the main infrastructure for the parsing of experiment output files in MATALB Mat-files are already implemented, and ii) the main statistical graphs are common for this class of agent-based models. However, programming skills are needed to implement model specific statistical graphs and measures. In the near future we plan to publish the Visualization Tool user manual and tutorial that will clearly describe tool functionality and the process of new models integration in the tool.
The models we outline here materialize the basic constituents of iCities approach on the formation and growth of Information Cities that emphasizes the role of increasing returns in explaining agglomeration and fast growth trends in the emerging World Wide Web economy. The results from the simulations we have obtained in this project clearly show that various self-reinforced feedback mechanisms (i.e. increasing returns) drive the organization of the web economy and explain a particularly strong pattern of concentration of people and economic activity in a few web locations.
In the last ten years increasing returns have been the subject of intensive research in economics. With increasing returns, economists say, competitive (i.e. shared) markets are not longer guaranteed, since if one product gets "ahead" in early stages of competition, it tends to stay ahead and ever increase its leading distance from the other products. The sources vary, but generally speaking, increasing returns operate: as large set up or fixed costs (giving an advantage to the large scale production units), learning benefits, coordination effects (giving an advantage to participations in formal or informal networks of agents taking similar action) and, of course, self-reinforcing expectations (usual behavior of the stock-markets). In this typology, iCities project adds a new category: informational increasing returns.
To describe how exactly web markets work under increasing returns, we have proposed two different but complementary views of the market-making functionality of the web environment.
1. According to the first view, Internet users interact each other or with web sites in many other ways than through the price mechanism. Or, putting it in a different way, web markets seem to have a strong institutional identity, in the sense that consumer choices (i.e. the choice to visit a web site) are simultaneously:
- Conditioned by individual tastes/preferences and,
- Embedded in identifiable action networks and ongoing relationships, which are created during the web navigation process. Recent studies on the behavior of the Internet users strongly support this view.
Two behavioral models based on this view. The starting question is the same: Why cities, information cities, i.e. large agglomerations of people and economic activity emerge in the virtual world? In the Internet, transportation costs are zero. Web sites can easily be reached from anybody and everywhere with no particular cost. In these conditions of equal access distance, one would rather expect a smooth web geography with a relatively even distribution of visitors per site. The first model we develop reproduces an interesting regularity of the web economy: the distribution of visitors per Internet site surprisingly follows a universal power law similar to what found in distribution of larger cities in US (and elsewhere) or in income distribution. The model provides a simple explanation of this phenomenon mostly based on the mechanisms through which information is transmitted in the Internet. The second model is actually a general class of models that capture the salient features of the dynamics of agglomeration in the web. The objective here is to examine evolutionary dynamics in the formation of the web geography and economy by studying an experimental environment, built as an agent-based computational model. We populate this environment with individual agents having preferences for web locations that depend upon locations attractiveness but we also assume benefits from agglomeration (when agents make choices about web sites, they receive a payoff depending on the number of agents having already visited that site at the time of choice).
2. The second view perceives the web markets as differentiated markets,
- Implying search costs for the consumer and,
- Imposing coalitions between web sites as a response to differentiated competition.
Participating in a coalition, this may be seen as a choice of a strategy in reaction to what other players are deciding. This is a typical non-cooperative game that seems to apply in many web cases, from Internet portal alliances to multi-product intermediaries and to collaborative commerce initiatives. A specific iCities model studies the formation and the evolution of such alliances.
3. Finally, we have build two additional models with particular interest as far as it concerns aggregation/segregation trends over the web. In the first model, we examine the influence these agglomerations may have on the successful diffusion of a certain category information products over the Internet. In the second model we outline principles of cluster formation in environments where users and suppliers locate in an information city while motivated from common preferences.
Recherche de données OpenAIRE...
Une erreur s’est produite lors de la recherche de données OpenAIRE
Aucun résultat disponible