Deep Q Networks for Classification: Performance Analysis on Iris and Diabetes Datasets

This study explores how Deep Q Networks (DQNS) can be applied to classifying data with standard datasets. Though DQNs are designed for sequential decisions, this study uses the Iris and Diabetes datasets, which are both used for classification, to test the performance. It investigates DQNs with different data imbalances and compares their results based on the quantity of data. Models are assessed using accuracy, precision, recall, F1 score, ROC-AUC, the amount of memory required, and their training speed. As demonstrated by the results, DQNs can match other algorithms in accuracy while attaining accuracy rates between 84% and 95% for various datasets and modified configurations. Based on the study, performance can be affected by adjusting hyperparameters and the distribution of data classes. DQNs are more complex than common classifiers such as SVMs and decision trees, and their main contribution in simple classification is yet to be proven. It adds to the enthusiasm for using reinforcement learning models in the context of supervised learning. The paper highlights the value of correct evaluation, points out risks linked to model over-fitting and includes new areas to pursue in the future such as benchmarking, clarifying models and using hybrid systems.