Deep Reinforcement Learning (DRL) provides a general optimization framework that has achieved strong results across many decision-making problems. In this study, we investigate the practical use of representative DRL algorithms as reward-optimized classifiers for galaxy morphology recognition on the Galaxy Zoo dataset. Each agent receives a galaxy image as observation, predicts a class label as an action, and obtains an immediate reward (+1 for correct, − 1 for incorrect); therefore, the learning problem corresponds to a single-step RL setting (contextual bandit/one-step MDP). To enhance robustness, we propose an ensemble approach using weighted majority voting, integrating predictions from multiple DRL agents into a multi-AI system. Our experiments on the Galaxy Zoo dataset shows that ensemble agent achieved highest accuracy of 98.55%, followed by DQN, RPPO with 98%, and PPO, A2C with 94%. Also, in order to evaluate the efficiency of DRL agents, a comparative study has been done with various deep learning models that are well established for galaxy image classification. We also report baseline results from standard deep-learning classifiers trained under the same data preprocessing and evaluation protocol. Our goal is to provide a transparent comparison and an accessible introduction to DRL-style optimization for image classification, while outlining sequential decision-making and continual/online learning extensions as future work
Yerramreddy et al. (Sun,) studied this question.