Calibration Strategies for Agent-based Simulation Models with Variability

Abstract

During the development of an agent-based simulation model, the model often has to be calibrated, which means adjusting the parameters such that a reference system can be reproduced. A major problem in calibrating an agent-based simulation model is the variability of the results, due to random choices made by the agents. To reduce the variability, the numbers of agents has to be increased, which in return increases the computation time of the simulation. An attempted solution to this problem consists of increasing the numbers of agents gradually. This approach is tested with two different calibration algorithm: simulated annealing and evolutionary algorithm. Different updating schedules are applied on a test model and examined in terms of their running time and their performance. It is shown that a evolutionary algorithm with an increasing agent count manages to produce similar results as a standard calibration using only half the computation time. To conclude, the best performing calibration process is used to calibrate an existing agent-based model simulating a well known past influenza epidemic.

Reproduction of patterns in melanocytic proliferations by agent-based simulation and geometric modeling

“…To align and compare our simulation results with data obtained with the standard visual diagnosis methods, we developed visual rendering techniques that display simulated melanocyte populations in dermatoscopic and histopathologic image styles. We used this visual approach to fine-tune the parameters of our simulation model in order to reproduce the quantitative characteristics of emerging macroscopic patterns and the temporal evolution of real nevi [ 62 ]. Beyond the two main patterns, reticular and globular, the morphologic appearance of acquired nevi may vary with regard to pigmentation, shape, size, and border abruptness [ 53 ].…”

“…For instance, the early phases in the evolution of skin lesions are rarely documented in clinical examination. Our approach could be of particular use to investigate this initial period by calibrating the model with intermediate to late phase sequential imagery and data [ 62 ]. Hypotheses constructed from our model could further serve as a starting point for future biomolecular studies.…”

Classification of Microscopic Models with Respect to Aggregated System Behaviour

“…More about these and other calibration methods can be found e.g. in [Rippinger et al, 2016,Rippinger, 2016. Though they are highly optimised and well-studied meta-heuristics, they all share the issue that successful calibration of a parameter value requires iterative execution of the simulation thousands of times.…”