If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at email@example.com.
Bayesian Network Structure Learning and its Application on Simulation-based Learning
AltmetricsView Usage Statistics
Bayesian network is a probabilistic graphical model that has wide applications in various domains due to its peculiarity of knowledge representation and reasoning under uncertainty. This research aims at Bayesian network structure learning and how the learned model can be used for reasoning. Learning the structure of Bayesian network from data is a challenging task because the number of possible structures increases super-exponentially with the increase in the number of variables. In order to address this challenge, we develop a method to efficiently learn the Bayesian network structure underlying the data set. Our novel hybrid Structure Learning Algorithm (hybrid-SLA) consists of two phases: the constraint-based phase restricts the search space and a score-and-search phase evolves the Bayesian network structure by exploring the reduced search space. In this dissertation, firstly, we investigate the effectiveness of using genetic algorithms as a score-and-search approach in our hybrid method. We evaluate the performance by comparing our proposed approach with the state-of-the-art algorithms over various well-known benchmark networks. The experimental results illustrate that our approach achieves good performance in inferring the most probable structure for the fully observed data. Secondly, we introduce a hybrid approach where we incorporate case-injected genetic algorithms as a score-and-search method that uses a caching mechanism for inducing Bayesian network structure. Our case-injected genetic algorithm enhance the performance over a sequence of similar problems by extracting and storing knowledge from previously solved problems and utilizing the accumulated knowledge in order to avoid repeated computations when solving subsequent related problems. To evaluate the viability of our approach, we conducted a series of experiments by generating a sequence of similar problems based on using data sets obtained randomly from various popular benchmark networks. Results show that our hybrid approach with case-injected genetic algorithm outperforms the genetic algorithm and another state-of-the-art algorithm in learning the Bayesian network structure. Finally, we apply our hybrid method to a simulation-based learning domain to perform structure discovery and to analyze the accuracy of the learned model in inferring the students’ difficult-to-learn concepts. Results show that the learned Bayesian network model is efficient in identifying such concepts.