The recent success of deep reinforcement learning in playing Atari games and Go revitalized reinforcement learning (RL) and provided inspiration to tackle real-world control problems based on deep neural perception. Some of the challenges in real-world applications of RL are the large sample complexity of RL methods, the need to define effective reward functions, and the potential risk (to the agent or the environment) in early stages of the learning process and when applying exploration actions. In this talk we discuss a training scheme, consisting of supervised elements, aiming to help in applying RL on real-world tasks. The supervised elements are imitation learning, reward induction, and safety module construction. We implemented this scheme using deep convolutional networks and applied it to create an agent capable of autonomous highway steering over the well-known racing game Assetto Corsa.
Hall: Hall D