A Quantitative Method for Comparing Multi-Agent-Based Simulations in Feature Space

Comparisons of simulation results (model-to-model approach) are important for examining the validity of simulation models. One of the factors preventing the widespread application of this approach is the lack of methods for comparing multi-agent-based simulation results. In order to expand the application area of the model-to-model approach, this paper introduces a quantitative method for comparing multi-agent-based simulation models that have the following properties: (1) time series data is regarded as a simulation result and (2) simulation results are different each time the model is used due to the effect of randomness, even though the parameter setups are all the same. To evaluate the effectiveness of the proposed method, we used it for the comparison of artificial stock market simulations using two different learning algorithms. We concluded that our method is useful for (1) investigating the difference in the trends of simulation results obtained from models using different learning algorithms; and (2) identifying reliable simulation results that are minimally influenced by the learning algorithms used.