A study on agent based evaluation methods to reduce damage at urban area

当灾害威胁到人类生命时，我们希望利用基于智能体的社会模拟（ABSS）来预测人的疏散行为和估计灾害的损害。当ABSS应用于实际应用时，有必要向潜在用户证明ABSS输出的有效性。在科学和工程领域，为了提高仿真的保真度，反复采用了猜测-计算结果-实验结果与仿真结果比较的过程。社会现象与可以用自然科学规律来解释的现象是相反的。社会现象与其说是客观衡量的，不如说是人类主观解释的。这给系统地分析社会现象和评价主体绩效带来了困难。在大多数情况下，很难获得真实案例的数据或进行实验来验证ABSS的结果和分析结果。本研究以灾害救援仿真系统为例，研究了不考虑应用领域相关指标的仿真输出分析方法，并给出了分析结果的解释。概率模型已被广泛用于模拟仿真参数和智能体特性的变化。利用大量的仿真数据对仿真结果进行了验证。例如，Johnson等人模拟了公共建筑的疏散，并提出了风险评估技术。我们的方法使用随机方法来表示代理的状态并分析它们的行为。分析方法是基于秩、特征值等数学量。它们独立于特定领域的评价，并表明该方法将应用于其他领域的潜力。

When disasters threaten human life, we want to use agent-based social simulation (ABSS) to predict the evacuation behavior of person and to estimate damage of disasters. When the ABSS is applied to practical usages, it is necessary to demonstrate the validity of the ABSS's outputs to the potential users. In scientific and engineering fields, the process Guess-Compute consequence-Compare experiment results with simulation results has been repeatedly used to increase the fidelity of simulations. Social phenomena are in contrast to the phenomena that can be explained using the laws of natural science.The Social phenomena are not so much objectively measured but subjectively interpreted by humans. It makes difficult to systematically analyze the social phenomena and to evaluate the performances of agents. In most cases, it is difficult to obtain data on real cases or conduct experiments to verify the results of ABSS and the analysis results.In this research, considering the disaster and rescue simulation system as an example of ABSS, methods to analyze the simulation outputs without any metrics related with the application domain have been studied and the interpretation of the analysis results are shown. The probability model has been widely used to model the variations in the parameters of simulation and agent characters. A large number of simulations with changing them statistically are used to verify simulation results. For example, Johnson, et al. have simulated the evacuation of public buildings and proposed risk assessment techniques. Our approach uses a stochastic approach to represent the states of the agents and analyze their behaviors.The analysis method is based on mathematical quantity such as rank, eigenvalue, etc. They are independent of domain specific evaluations and suggest the potential of the method that will be applied to other fields.