Human Activity Recognition (HAR) is an important task in applications such as healthcare monitoring, wearable devices, IoT systems and smart environments. This includes recognising human activities using sensor data such as accelerometers and gyroscopes. Traditional machine learning and deep learning models such as SVM, CNN, RNN, and LSTM have demonstrated good performance but face challenges such as imbalanced datasets, poor recognition of rare activities, and inability to adapt to dynamic changes in human behaviour.This work proposes a novel approach called Dynamic Reinforcement Learning-Based GOWLA (RL-GOWLA) to overcome these limitations. The method combines LSTM networks for temporal feature extraction with a Weighted Logarithmic Averaging (GOWLA) technique for prediction aggregation. A reinforcement learning agent is introduced to adaptively adjust aggregation weights according to the performance feedback, leading to better adaptability and classification accuracy.The experimental results on the WISDM dataset demonstrate that the proposed model achieves a high F1-score of 96.58% and outperforms traditional approaches. The system effectively improves the recognition of rare activities, reduces overfitting and adapts to temporal variations. In general, the proposed RL-GOWLA framework is a robust, scalable and efficient solution for real-world HAR applications in healthcare, IoT, and smart monitoring systems.
KUMAR et al. (Fri,) studied this question.