Show only achievement: Implement training autopilot A3C-end model based TORCS and Torch7
TORCS based blogger and Torch7 developed based on continuous operation of the autopilot to-end simulation system depth reinforcement learning models, is a multi-threaded asynchronous training methods, each thread updated summary of the total gradient model parameter update, then the overall model parameter update issued minutes after the child thread, the child thread continue to update on this basis, in fact, A3C. Model input is a front perspective image of the vehicle, i.e. of the main viewing screen shot simulation interface, the output value of the rotation angle is a normal probability distribution.
Bloggers are based on existing projects others changed, the project has to do is discrete actions , I've turned into a continuous motion , while repairing some hidden bug. And bloggers use this only as a validation of the proposed own pre-training algorithm simulation platform with the bloggers themselves but also with ROS a distributed real vehicle training system based on three real car.
Because involve their own research, the need for confidentiality, previous studies it is also not disclosed.
