Abstract
Models of social systems generally contain free parameters that cannot be evaluated directly from data. A calibration phase is therefore necessary to assess the capacity of the model to produce the expected dynamics. However, despite the high computational cost of this calibration it doesn't produce a global picture of the relationship between the parameter space and the behaviour space of the model. The Calibration Profile (CP) algorithm is an innovative method extending the concept of automated calibration processes. It computes a profile that depicts the effect of each single parameter on the model behaviour, independently from the others. A 2-dimensional graph is thus produced exposing the impact of the parameter under study on the capacity of the model to produce expected dynamics. The first part of this paper is devoted to the formal description of the CP algorithm. In the second part,we apply it to an agent based geographical model (SimpopLocal). The analysis of the results brings to light novel insights on the model.
Highlights
1.1 Agent-based modelling is widely used in models of social systems because it is well suited to describe individual-centred dynamics with non-linear interactions
9.1 This paper proposes a new method to explore models. This method, called Calibration Profile, computes 2-dimensional graphs that depict the effect of each parameter on the model dynamics using a calibration objective
It provides a global view of the local effect of each parameter on the expected behaviour. This constitutes a new form of sensitivity analysis of parameters on a calibration objective that differs from classical sensitivity analysis methods (Saltelli et al 2000)
Summary
1.1 Agent-based modelling is widely used in models of social systems because it is well suited to describe individual-centred dynamics with non-linear interactions. 1.2 The current prevalent method for automatic calibration is to see it as an optimisation problem in which the differences between known data and model predictions have to be minimised This optimisation can be performed with different optimisation algorithms like "Approximate Bayesian Computation" (Lenormand et al 2012) or evolutionary algorithms (Schmitt et al 2014; Stonedahl 2011). While classic stochastic optimization algorithms search for a single optimum of a function, the CP method aims to fill a 1-dimensional array of "categories" (a 2-D map for MOLE) that should each contain the best performing sample for each possible value of the parameter under scrutinity (after a discretization). The projection of the last selected points along x1 constitutes an approximation of the theoretical continuous profile (step 4)
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call