This paper explores the application of deep learning, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) for Human Activity Recognition (HAR) using wearable sensors. The authors rigorously evaluate these models across three representative datasets: Opportunity, PAMAP2, and Daphnet Gait, which cover a range of activities from manipulative gestures to physical exercises and medical conditions like Parkinson's disease. They describe the training process for deep, convolutional, and recurrent models, introduce a novel regularization approach for RNNs, and investigate the suitability of each model for different tasks in HAR. Through over 4,000 experiments, they analyze the impact of hyperparameters using the fANOVA framework and provide guidelines for practitioners. The results show that bi-directional LSTMs outperform state-of-the-art methods on the Opportunity dataset, and recurrent networks generally outperform CNNs for activities with natural ordering and short durations. The study highlights the importance of parameter exploration and provides insights into the performance variability of different models, suggesting that practitioners should focus on tuning learning parameters first.This paper explores the application of deep learning, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) for Human Activity Recognition (HAR) using wearable sensors. The authors rigorously evaluate these models across three representative datasets: Opportunity, PAMAP2, and Daphnet Gait, which cover a range of activities from manipulative gestures to physical exercises and medical conditions like Parkinson's disease. They describe the training process for deep, convolutional, and recurrent models, introduce a novel regularization approach for RNNs, and investigate the suitability of each model for different tasks in HAR. Through over 4,000 experiments, they analyze the impact of hyperparameters using the fANOVA framework and provide guidelines for practitioners. The results show that bi-directional LSTMs outperform state-of-the-art methods on the Opportunity dataset, and recurrent networks generally outperform CNNs for activities with natural ordering and short durations. The study highlights the importance of parameter exploration and provides insights into the performance variability of different models, suggesting that practitioners should focus on tuning learning parameters first.