Adaptive Reinforcement Learning Agent for Orchestrating Extreme Learning Machine and Transfer Learning in Flower Classification
, Priyanka Hanumanthappa, Deepa Shetty, Prthu Rao HThe purpose of this study is to develop an adaptive real-time flower classification framework that maximizes accuracy, speed, and resource efficiency by integrating Transfer Learning, Extreme Learning Machines, and Reinforcement Learning.
Accurate and efficient flower species identification is essential for applications in smart city planning, agriculture, and floriculture. Traditional static classification models lack adaptability to diverse environments and resource constraints, necessitating dynamic approaches that can optimize performance in real-time settings.
The proposed framework employs a Reinforcement Learning (RL) agent to dynamically select between lightweight classifiers (ELM + MobileNetV2) and high-accuracy classifiers (CNN + EfficientNetB0). The system is trained and evaluated on a public Kaggle dataset and real-time images captured via IoT-enabled cameras, ensuring robustness in both controlled and real-world scenarios.
This study introduces a novel adaptive classification system that balances accuracy, inference latency, and resource usage through RL-based decision-making. It advances the field by demonstrating superior adaptability and efficiency compared to static ensembles and state-of-the-art methods.
The framework achieves 95.8% accuracy while reducing training and inference latency. The RL-driven approach outperforms traditional static models, showing enhanced scalability, resource-awareness, and real-time performance in diverse environments.
Practitioners should consider implementing RL-based adaptive classification systems to improve real-time accuracy and resource management, especially in IoT-driven environments, such as smart cities and precision agriculture.
Further research is encouraged to explore additional classifier combinations, extend the framework to other classification tasks, and investigate long-term adaptive strategies for dynamic environments.
This framework has the potential to significantly improve real-time monitoring and management in urban greenery, crop monitoring, and flower industry automation, contributing to sustainable practices and enhanced urban aesthetics.
Future studies may focus on integrating more diverse classifiers, optimizing RL policies for even faster adaptation, and deploying the system in large-scale, real-world deployments to evaluate long-term robustness and utility.
Back