Nuevos Métodos basados en Soft Computing para Calibración de Modelos Basados en Agentes: Aplicaciones en Marketing y Ciencias Políticas

Abstract

Esta tesis doctoral trata el problema de calibrar y validar modelos basados en agentes (ABMs). En concreto, se propone mejorar las técnicas de calibración existentes utilizando metaheurísticas novedosas y proponiendo un framework integral para la calibración de modelos con múltiples indicadores de rendimiento (KPIs). Los avances propuestos se aplicaron a modelos para escenarios de ciencias políticas y de marketing.

Entre las principales contribuciones podemos destacar el diseño de un motor de simulación de ABMs para escenarios de ciencias políticas y marketing, con el que se compuso un framework de calibración para ABMs. También destacamos el diseño y validación de distintos algoritmos de calibración para uno o más KPIs basados en metaheurísticas novedosas. Finalmente, destacamos también la propuesta de un framework integral para calibración y validación que combina un algoritmo EMO con una técnica de visualización avanzada basada en grafos.

Model simulation is an established approach for analyzing complex systems but these computational models need to be carefully calibrated and validated before they can be useful. This task can be challenging in many cases as the lack of information prevents the users for precisely estimating the parameters of the model and the number of parameters to estimate can be too high. The agentbased model methodology is a well-known model simulation approach that relies in the behaviour of artificial agents, which are autonomous entities that act following simple rules and interacting with other agents and their environment. Due to the mentioned issues, the calibration of agent-based models is usually carried out using automated calibration methods, since they can estimate those parameters which cannot be set because of the lack of information. However, the use of automatic calibration does not release the user from carefully reviewing and validating the resulting parameter set, as a good fitting between the models' output and the calibration data is not a guarantee of a valid configuration. This doctoral dissertation addresses these issues in several ways. First, it proposes multiple agent-based models for defining political scenarios and marketing strategies that serve as the foundation of the dissertation, since they are calibrated and validated with the techniques and methods proposed by this thesis. In addition, these agent-based models are useful for obtaining insights relevant to their application domain, which is two-fold. On the one hand, two models tackle the Spanish national elections on the 14th of March of 2004, that were severely in uenced by the terrorist attacks that happened three days before. On the other hand, we introduce a benchmark containing several instances of a model for defining marketing strategies that considers the awareness of the brands and their word-of-mouth volume. This doctoral dissertation reviews and compares several relevant metaheuristics to design the best performing method for agent-based model automatic calibration. The use of a well performing optimization method is important for automatic calibration as its success depends on the method's ability for exploring the complex and ill-defined parameter search space. Thus, this dissertation conducts an exhaustive experimentation comparing well-established and recent evolutionary algorithms and including their hybridization with local search procedures. The computational study analyzes the calibration accuracy of the metaheuristics using an integer coding scheme over the developed benchmark of instances of the agent-based model for marketing, which considers an increasing number of decision variables. This study reported the outstanding performance of the memetic coral reefs optimization algorithm after performing multiple statistical tests to the results. The choice of the best performing metaheuristic is specially relevant for models calibrated over more than a single output, as its calibration requires handling different criteria jointly. This fact increases the problem complexity and can be achieved by using automated calibration and evolutionary multiobjective optimization methods as they can find a set of representative Pareto solutions under these restrictions and in a single run. However, selecting the best algorithm for performing automated calibration can be overwhelming. Therefore, this dissertation proposes an exhaustive analysis of the performance of several evolutionary multiobjective optimization algorithms when calibrating different instances from the agent-based model for marketing. This analysis evaluates the performance of the compared algorithms by using multiobjective performance indicators and attainment surfaces, including a statistical test for studying the significance of the indicator values, and benchmarking their performance with respect to a classical mathematical optimization method. The results of this experimentation re ect that those algorithms based on decomposition perform significantly better than the remaining methods in most instances. Moreover, we also identify how different properties of the problem instances (i.e., the shape of the feasible search space region, the shape of the Pareto front, and the increased dimensionality) erode the behavior of the algorithms to different degrees. Finally, this dissertation introduces a multicriteria integral framework to assist the modeler in the calibration and validation processes of agent-based models that combines evolutionary multiobjective optimization with network-based visualization. Up to our knowledge is the first integral approach to model calibration in the specialized literature. This approach combines the outstanding performance of evolutionary multiobjective optimization algorithms with an advanced visualization method to better understand the decision space and the set of solutions from the obtained Pareto set approximation. This proposal is tested by using two instances of the agentbased model for marketing. The final analysis of the calibrated solutions shows how the proposed framework eases the analysis of Pareto sets with high cardinality and helps with the identification of exible solutions (i.e., those having close values in the design space), thus helping the designer in the agent-based model validation.