Complexity and Performance Evaluation of Segmented and Recursive Reinforcement Learning

Artificial Intelligent systems play very crucial role in current advanced technology with complex Machine Learning and Deep Learning algorithms. Reinforcement Machine Learning models directly learn from the environment and very useful for solving many complex problems in absence of prior datasets. Existing reinforcement algorithms like Q-learning solve complex problems in game theory and data science with high accuracy. But most of the reinforcement learning algorithms store the reward, parameters and state-data at each state, require huge storage and computational resources and results in large training periods. Lower training period is very important in real-time continual and transfer learning systems to reduce the latency. In this paper, novel Segmented and Recursive Reinforcement Learning algorithm is proposed to meet the time and resource constraints of the intelligent systems with very less training period. The performance of proposed algorithm is evaluated by comparing it with the existing Q-learning and Markov decision Process based algorithms in terms of accuracy, complexity and training periods. Both proposed and existing reinforcement algorithms are used for Parameter tuning in wavelet based image noise reduction application and accuracy of the algorithms are analyzed in terms of the noise reduction performance.