Boosting Weak Learners With Multi-Agent Reinforcement Learning for Enhanced Stacking Models: An Application on Driver Emotion Classification

Recently, there has been an increasing interest in the effects of stress and negative emotions on driving performance and road safety. As a consequence, there is a growing interest in studies that employ biometric signals to categorize the emotions of drivers, and driver state monitoring technologies are assuming a greater level of significance within the automotive sector. The objective of this study is to develop a lightweight stacking model that classifies drivers’ emotions into seven distinct categories by combining statistical electroencephalography data, psychological survey data, and driver behavior data. Our objective is to effectively combine individual machine learning models using reinforcement learning, and achieve optimal performance by combining strong and weak learners. To overcome the drawbacks of previous works that select individual models arbitrarily and optimize the weight in assemble model, we propose a multi-agent reinforcement learning based model selection for stacking. The proposed model introduces a novel feature by providing varying rewards based on the contribution of individual model to the overall performance, with the aim of enhancing the weaker model selection. The final results show that the meta-model achieves the highest performance when using a decision tree, with an accuracy of 0.8543 and an F1-score of 0.8462, which is an improvement of 0.266 and 0.2875 compared to before applying reinforcement learning, respectively. Experimental findings validate that our method can contribute to the improvement of road safety through the identification of emotional shifts in drivers and the development of appropriate interventions.