Exploiting Reward Machines with Deep Reinforcement Learning in Continuous Action Domains

Deep reinforcement learning can solve real-world robot control problems, such as autonomous driving and robotic arm manipulation. In deep reinforcement learning, an agent does not know the problem description and learns the optimal solution through trial-and-error. This method brings two major challenges when solving real-world problems: partial observability and learning efficiency. In this thesis, we address these two challenges and extend previous work. First, we use reward machines to address the problem of partial observability. Then, we focus on finding the existing cutting-edge deep reinforcement learning algorithms and integrating them with reward machines to enhance the learning efficiency. To test the performance of all the algorithms, we proposed a series of different tasks that can be used to mimic real-world robot control problems. Finally, based on the test results, we compare the performance of all the algorithms and analyze their advantages and disadvantages.