Agent-based Simulations using Data-Oriented Random Numbers

M. Hosoi and Y. Uchida (Japan)


Agent-based Simulation, Multi-Agent Systems, Computa tional Economics, Computational Statistics


In agent-based simulations in the field of economics or po litical science, when the probability distribution of a key parameter is unknown and/or multiple systems or alterna tive policy can be suggested to fit the results, it is desir able to perform a statistical assessment to see which sys tem or countermeasure is most likely to be effective. How ever, no general procedure has been established for do ing so. Therefore we have been advocating the Numeri cal Probability Distribution as a method for evaluating the results of simulations and testing hypotheses and for pre pare data-oriented random numbers when the variables of the simulations and/or simulation results have distributions that are unknown. This paper presents two agent-based simulations, in which one was performed with pseudodata sets prepared with correlations matching the correlation ob served between two of its variables, and the other was per formed with pseudodata sets that were completely indepen dent. The results of the two simulations were compared and evaluated using the NPD method, and this analysis in dicated that there is a potential for simulations to lead to in correct conclusions if their pseudodata does not account for actual correlations between variables. This paper demon strated how useful it is to employ the NPD method in agent based simulations.

Important Links:

Go Back