Deep Reinforcement Learning with Demonstration on the OpenAI Environment, Car Racing-V0 скачать в хорошем качестве

Deep Reinforcement Learning with Demonstration on the OpenAI Environment, Car Racing-V0

In this project, a DQN agent is implemented with the method of deep reinforcement learning with demonstration in the OpenAI Gym environment called Car Racing-V0. The main idea behind this approach is to provide the agent with the most useful observations collected by a human expert, rather than allowing the agent to explore the environment on its own and collect insignificant experiences with the pure reinforcement learning methods. The environment, Car Racing-V0, has a continuous action-space. There are three types of actions in this environment, namely steering wheel, acceleration, and brake [steering wheel, acceleration, brake]. Steering wheel has the value between -1 and +1, whereas acceleration and brake have the values between 0 and +1. For simplicity, action space is discretized so that there are only five actions that the agent can take, which are turning right [+0.25, 0.0, 0.0], turning left [-0.25, 0.0, 0.0], soft acceleration [0.0, 0.5, 0.0], soft deceleration [0.0, 0.0, 0.4], and doing nothing [0.0, 0.0, 0.0]. Default observation space of this environment is a an RGB image with the shape of (96x96x3). However, to simplify the state space, the lower panel is removed and the image is converted from rgb to gray scale. Finally, the frame is shrunk to the size of (70x70x1). To provide the speed information to the agent, 4 consecutive frames are passed to agent as a single state.