A Robust Approach for Continuous Interactive Actor-Critic Algorithms

Reinforcement learning refers to a machine paradigm in which an agent interacts with the environment learn how perform task. The characteristics of may change over time or be affected by disturbances not controlled, avoiding finding proper policy. Some approaches attempt address these problems, as interactive reinforcement learning, where external entity helps through advice. Other approaches, such robust allow task, acting disturbed environment. In this paper, we propose approach that addresses problems dynamic environment, advice provides information on task and dynamics Thus, learns policy while receiving We implement our version cart-pole balancing simulated robotic arm organize objects. Our results show proposed allows complete satisfactorily dynamic, continuous state-action domain. Moreover, experimental suggest agents trained are less sensitive changes than agents.