Discriminative Parameter Learning of Bayesian Networks Using Differential Evolution: A Preliminary Analysis Alejandro Platas L´ opez, Nicandro Cruz Ram´ ırez, Efr´ en Mezura Montes, Alejandro Guerra Hern´ andez Universidad Veracruzana, Centro de Investigaci´ on en Inteligencia Artificial, M´ exico alejandroplatasl@gmail.com,{ncruz,emezura,aguerra}@uv.mx Abstract. This work proposes Differential Evolution (DE) to train pa- rameters of Bayesian Networks (BN) for optimizing the Conditional Log-Likelihood (Discriminative Learning) instead of the log-likelihood (Generative Learning). Although Discriminative Parameter Learning al- gorithms have been proposed, to the best of the authors’ knowledge, a metaheuristic approach has not been devised yet. Thus, the objective of this research is to come up with this kind of solution and evaluate its behavior so that its feasibility in this domain can be determined. According to the theory such a solution tends to generate low-bias classi- fiers that minimize classification error but this is not reflected in results, regarding proposed method, bias in search for best solutions improves DEs performance. Keywords: Bayesian networks, differential evolution, discriminative pa- rameter learning. 1 Introduction Two paradigms are distinguished for parameter learning of Bayesian networks. One of them, called Generative Learning (GL), optimizes Log-Likelihood in order to obtain the parameters that characterize the joint distribution in the form of local conditional distributions, and subsequently estimates class conditional probabilities using the Bayes rule. Even though this paradigm is computationally efficient, it is likely to generate biased classifiers [12]. The other paradigm optimizes Conditional Log-Likelihood (CLL) to directly estimate the parameters associated with conditional class distribution. Such paradigm is known as Discriminative Learning (DL) and generates low-bias classifiers that typically tend to minimize the classification error. In addition, the effect caused by the assumption of conditional independence among attributes in the network structure, but which may be violated in the data, is reduced. 75 ISSN 1870-4069 Research in Computing Science 149(3), 2020 pp. 75–82; rec. 2019-09-17; acc. 2019-09-30