Agent-based modeling (ABM) is emerging as a powerful tool for the analysis of human behavior in several situations in which direct observation is impossible or unfeasible [4]. This computational simulation technique is still at its early stages in the social sciences [24], with sociology starting before other disciplines [7] and management reluctantly catching up [19]. The enthusiasm that many have shown with ABM is witnessed by the increasing number of scientific outlets that publish such simulations-among many are the journal of Computational and Mathematical Organization Theory and the Journal of Artificial Societies and Social Simulation. It has been pointed out that the diffusion of ABM is to be found, among other aspects, in an increase of available computational power [6], a set of simplified tools for modeling [19], and the ability to create extremely complex and descriptive simulations [5]. As a result, these simulations are usually employed to study configurations, structures, and outcomes that are not directly intelligible by a simple review of the initial conditions of a system. In other words, ABM are a tool to study emergent properties of complex systems among other characteristics. In spite of the very significant use of these simulations, there is still uncertainty on the conditions under which a simulated system produces relevant or irrelevant results-in relation to a given research question. The issue is particularly relevant for the study of emergent properties in that some of them may result from misinterpreting the data of the simulated model. These issues can be divided into two, separate but (probably) interconnected.We write 'probably' because, to our knowledge, there are no studies robustly and soundly connecting these two issues together. The first issue relates to the question: when is it the right time to stop the simulation in a single run? This is a problem concerning the length of a run-some software may call these 'steps'-, given a certain configuration of the simulation's parameters. It can be rephrased as to when it is enough time for emergent properties to emerge. Length may depend on the research question, on the type of simulation, or it can be calculated [20]. To partially address these issues, some propose steady state and stationarity analyses [8]. The second issue relates to the number of runs per configuration of parameters: how many times should one run the simulation for each configuration of parameters? This is a problem of establishing an appropriate number such that the data is not biased or that effects are discarded when they should not. Some have suggested power analysis as a viable technique to tackle with this problem [16, 21, 22]. Given the relevance of this latter issue, this abstract is dedicated to it, in an attempt to link emergent properties of the ABM to the appropriate number of runs to be performed.
Power analysis for agent-based modeling determining the appropriate number of runs
SECCHI, DAVIDE;Seri, Raffaello
2017-01-01
Abstract
Agent-based modeling (ABM) is emerging as a powerful tool for the analysis of human behavior in several situations in which direct observation is impossible or unfeasible [4]. This computational simulation technique is still at its early stages in the social sciences [24], with sociology starting before other disciplines [7] and management reluctantly catching up [19]. The enthusiasm that many have shown with ABM is witnessed by the increasing number of scientific outlets that publish such simulations-among many are the journal of Computational and Mathematical Organization Theory and the Journal of Artificial Societies and Social Simulation. It has been pointed out that the diffusion of ABM is to be found, among other aspects, in an increase of available computational power [6], a set of simplified tools for modeling [19], and the ability to create extremely complex and descriptive simulations [5]. As a result, these simulations are usually employed to study configurations, structures, and outcomes that are not directly intelligible by a simple review of the initial conditions of a system. In other words, ABM are a tool to study emergent properties of complex systems among other characteristics. In spite of the very significant use of these simulations, there is still uncertainty on the conditions under which a simulated system produces relevant or irrelevant results-in relation to a given research question. The issue is particularly relevant for the study of emergent properties in that some of them may result from misinterpreting the data of the simulated model. These issues can be divided into two, separate but (probably) interconnected.We write 'probably' because, to our knowledge, there are no studies robustly and soundly connecting these two issues together. The first issue relates to the question: when is it the right time to stop the simulation in a single run? This is a problem concerning the length of a run-some software may call these 'steps'-, given a certain configuration of the simulation's parameters. It can be rephrased as to when it is enough time for emergent properties to emerge. Length may depend on the research question, on the type of simulation, or it can be calculated [20]. To partially address these issues, some propose steady state and stationarity analyses [8]. The second issue relates to the number of runs per configuration of parameters: how many times should one run the simulation for each configuration of parameters? This is a problem of establishing an appropriate number such that the data is not biased or that effects are discarded when they should not. Some have suggested power analysis as a viable technique to tackle with this problem [16, 21, 22]. Given the relevance of this latter issue, this abstract is dedicated to it, in an attempt to link emergent properties of the ABM to the appropriate number of runs to be performed.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
