Abstract
Reinforcement learning is well known for its ability to model sequentialtasks and learn latent data patterns adaptively. Deep learning models have beenwidely explored and adopted in regression and classification tasks. However,deep learning has its limitations such as the assumption of equally spaced andordered data, and the lack of ability to incorporate graph structure in termsof time-series prediction. Graphical neural network (GNN) has the ability toovercome these challenges and capture the temporal dependencies in time-seriesdata. In this study, we propose a novel approach for predicting time-seriesdata using GNN and monitoring with Reinforcement Learning (RL). GNNs are ableto explicitly incorporate the graph structure of the data into the model,allowing them to capture temporal dependencies in a more natural way. Thisapproach allows for more accurate predictions in complex temporal structures,such as those found in healthcare, traffic and weather forecasting. We alsofine-tune our GraphRL model using a Bayesian optimisation technique to furtherimprove performance. The proposed framework outperforms the baseline models intime-series forecasting and monitoring. The contributions of this study includethe introduction of a novel GraphRL framework for time-series prediction andthe demonstration of the effectiveness of GNNs in comparison to traditionaldeep learning models such as RNNs and LSTMs. Overall, this study demonstratesthe potential of GraphRL in providing accurate and efficient predictions indynamic RL environments.