Exploring Reinforcement Learning Algorithms for Autonomous Agents in Dynamic Environments

Reinforcement Learning (RL) algorithms are widely considered to be the enabling technology for deep learning and allow for agents in complex systems to learn in a dynamic environment. In order to present in this paper a break-through training methodology for learning RL agents in the dynamic environment, advised by the following adaptability, robustness, and scalability purposes. Our approach is assessed in an array of the model environments which are open to the users. This framework is labelled as OpenAI Gym environment. We apply different reinforcement learning algorithms in our experiments among which Covering Concurrent and Differentiable in Q-learning, Meta-learning (MAML), Virtual Imitation and Q-learning, and Dyna-Q are the most popular. Meta-learning’s performance highlights its power and efficiency in grasping and optimizing decision-making logic in dynamic situations within 20 epochs and reaching the cumulative reward of 500. As a matter of fact, adaptive search strategies, for instance, UCB are another tool that is extremely useful reaching to 490 at 25 epochs after all. Transfer learning techniques like the Progressive Neural Networks shows extremely high power of learning in collecting a reward of 500 and in less than 20 epochs it also converges. Moreover, scalable parallel training methods such as CPDG which are normalized in distribution, have huge scalability and robustness can be applied for large scale dynamic environments. Consequently, it is concluded that the presented approach was very good in giving RL agents the ability to navigate environments and succeed in dynamic settings, thus contributing to the development of autonomous systems not only in the area of RL but also everywhere else.