Belief rule-based expert systems (BRBESs) are widely used in various domains which provide an integrated framework to handle qualitative and quantitative data by addressing several kinds of uncertainty. The correctness of the data significantly affects the accuracy of the BRBESs. Learning plays an important role in BRBESs to upgrade their knowledge base and parameters values, necessary to improve the accuracy of prediction. In addition, comparatively larger datasets hinder the accuracy of BRBESs.

Therefore, this doctoral thesis focuses on four different aspects of BRBESs, namely, the accuracy of data, multi-level complex problem, learning of BRBES, and accuracy of prediction for comparatively large dataset.

First, the accuracy of data acquisition plays an important role, necessary to ensure accurate prediction in BRBESs. Therefore, the data coming from sensors contain anomaly due to various types of uncertainty, which hampers the accuracy of prediction. Hence, anomalous data needs to be filtered out. A novel algorithm based on belief rule base for detecting the anomaly from sensor data has been proposed in this thesis.

Second, BRBESs can be considered to handle the multi-level complex problem like the prediction of a flood as they address different types of uncertainty. A web based BRBES was developed for predicting flood which provides better usability, allows handling of larger numbers of rule bases, and facilitates scalability. In addition, a learning mechanism for multi-level BRBESs has been developed by taking account of flooding, considered as an example of a complex problem. This learning mechanism for multi-level BRBES demonstrates promising results in comparison to other machine learning techniques including, Long Short-term Memory (LSTM), Artificial neural network (ANN), Support Vector Machine (SVM), and Linear regression.

Third, different optimal training procedures used to support learning in BRBESs. Among these, Differential Evolution (DE) appears performing better in comparison to other evolution algorithms, including Particle Swarm Optimisation (PSO) and Genetic Algorithm (GA). However, DE's performance depends considerably in assigning near-optimal values to its control parameters. Therefore, an enhanced belief rule-based adaptive differential evolution (eBRBaDE) proposed in this thesis with the capability of ensuring balanced exploitation and exploration in the search space by providing near-optimal values to the DE's control parameters. The capability of accurate prediction of eBRBaDE has been demonstrated by taking account of power usage effectiveness (PUE) of datacentre in comparison to other evolutionary algorithms used in BRBESs optimal training procedures.

Fourth, the recent advancement of sensor technologies enabled acquiring of a huge amount of data. In this context, deep learning appears as an effective method to process this huge amount of data. However, this high volume of data contains various types of uncertainties, including vagueness, imprecision, randomness, ignorance and incompleteness. Hence, an enhanced deep learning approach, named BRB-DL, has been developed by integrating BRBES, allowing the improvement of prediction accuracy, especially in case of a large dataset. The applicability of this BRB-DL has been carried out by considering a large amount of air pollution data to predict the air quality index (AQI) of different Chinese cities.

In the light of the above, it can be argued that the novel anomaly detection algorithm proposed in this thesis enables the removing of anomalous data. The proposed learning mechanism for multi-level BRBES allows handling of the multi-level complex problem. The optimal training procedure, named eBRBaDE, enabling determination of optimal learning parameters of BRBESs and finally, the integration of deep learning with BRBES allows to handle large data set.