Research Area:  Machine Learning
In this work, an efficient and simple image recognition classification system has been proposed. It consists of components from both reinforcement learning and deep learning. More specifically, Q-learning is used with an agent having two states, and two to three actions. This classifier is different from others, because the latter uses features of convolutional nets and also uses past histories in addition to Q-states. The other techniques found in literature have issues due to the large number of states used, as the dimensions of their feature maps are quite large. Since the novel technique proposed as only two Q-states, it has the advantage of being simple and also having significantly lesser parameters to optimize. Also, it has a straightforward reward process. Another advantage of the proposed classifier is usage of unique action set for image processing not found in literature. Accuracy of the proposed classifier has been compared with various contemporary algorithms on important datasets from ImageNet, Caltech-101, and Cats and Dogs . The classifier given in this work performs better than other classifiers on the various datasets used experimentally.
Keywords:  
Image Classification
Reinforcement Learning
Two-State Q-Learning
Q-learning
Optimization
Author(s) Name:   Abdul Mueed Hafiz
Journal name:  Handbook of Intelligent Computing and Optimization for Sustainable Development
Conferrence name:  
Publisher name:  Wiley Online Library
DOI:  10.1002/9781119792642.ch9
Volume Information:  
Paper Link:   https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119792642.ch9