A solid waste classification using reinforcement learning boosted by data augmentation and hyperparameter optimization.

With the rising rate of urbanization and the high rate of development of smart cities, handling the increasing amount of waste, both in terms of their production, sorting, and discarding, has become an urgent task. New plans have used the convolutional neural network (CNN) techniques, and huge pre-labeled image data to enhance the control of garbage including sorting, dump, disposal and recycling. Nonetheless, formation and labeling of such datasets is also time and cost demanding. In order to solve this problem, the proposed study aims to use an active learning (AL) framework that would enhance categorization of waste types, including organic, and recyclable, with less labelled examples. Normal models of AL tend to apply fixed selection mechanisms, which are not able to be adjusted to evolving data. The model combine the deep reinforcement learning (DRL) with an innovative scope loss function (SLF) to improve adaptability. The purpose of this step is to balance the trade-off between exploration and exploitation of new information. Besides, online data augmentation is performed with the help of a generative adversarial network (GAN). An additional regularization approach is incorporated in order to make GANs training more stable and reduce the likelihood of mode collapse. Moreover, a better algorithm of hyperparameters tuning (pattern) is applied and a k-means mutation technique is employed to select the most suitable ones. The efficiency of the proposed method is proved on three benchmark datasets, namely TrashNet, Trash, and OrgalidWaste. The active learning process starts with only 10% labeled samples, while the remaining 90% are treated as unlabeled. At each iteration, the DRL agent selects a small batch of informative samples under a fixed annotation budget. The classifier is retrained incrementally after each selection round until the labeling budget is exhausted. F-metrics of the model are 92.29, 89.27 and 89.90, indicating distinct advancements in waste classification using images and few labelled data. Generally, this research would help in creating a more intelligent and efficient wastes management systems.