<p>Facial Expression Recognition (FER) is currently a very active field of research. It involves a computer’s capability to recognize and interpret human emotional expressions, which change with an individual’s internal emotional state. Several researchers have been working on this topic, using classical methods or Neural Network (NN) approaches. The accuracy and efficiency of classification models are highly dependent on the quality of the training data. The training process becomes much more challenging when there is a lot of redundant, irrelevant, noisy, or unreliable data. Although many techniques have been developed to address data quality problems; however, these have not been entirely effective. To tackle this critical issue, we propose a Reinforcement Learning with a Genetic Algorithm (RLGA-FER) approach to enhance the quality of training data for FER systems. The proposed RLGA-FER system consists of two components: an agent and an environment. The agent uses a genetic algorithm to make a global decision about whether an image should be retained in the dataset or removed. By leveraging reinforcement learning, the agent dynamically learns and adapts to select the most relevant images, while the genetic algorithm robustly explores and evolves the selection process for optimal performance. The environment consists of three parts: a global training dataset, a feature extractor, and a recognition system. The aim of the feature extractor is to generate feature vectors for the dataset by applying a Conventional Neural Network (CNN). The retained feature data are used to train the evaluation system, and a Support Vector Machine (SVM) is used for evaluation. Our RLGA-FER system is evaluated on two popular FER datasets, RAF-DB and ExpW. The experimental results demonstrate that the proposed RLGA-FER system performs well, with recognition rates of 85.20%, and 77.34% for the RAF-DB, and ExpW datasets, respectively.</p>