Training system for autonomous driving control policy

A training system for an autonomous driving control policy, comprising three modules of a construction of a simulator, a policy search, and a policy transfer;

• the construction of the simulator:
• a simulation to static factors such as power systems of vehicles and driving roads as well as a simulation to dynamic factors such as pedestrians, non-motor vehicles, and surrounding vehicles are involved;
• the policy search:
• in the simulator which is constructed, a driving objective function is set, and a driving control policy of an optimal objective function is searched for by means of a machine learning method;
• and
• the policy transfer:
• the policy searched out in the simulator is retrained according to data acquired by an unmanned vehicle entity to obtain a driving control policy used for the unmanned vehicle entity, wherein
• the policy transfer comprises:
• initialization of a transfer model:
• a control policy model is run in an unmanned vehicle entity with an autonomous driving control policy model obtained through training in the simulator as a starting point, and is updated by means of obtained data;
• a simulator transition correction and transfer:
• firstly, a control action sequence (a1, a2, a3, . . . , an) is executed on the unmanned vehicle entity, and perception states (s0, s1, s2, s3, . . . , sn) of all executed action are collected;
• secondly, in the simulator, an initial state is set as s0, and a same action sequence (a1, a2, a3, . . . , an) is executed;
• and perception states (s0, u1, u2, u3, . . . , un) are collected;
• thirdly, in the simulator, a transition correction function g is constructed, an action a from a current state s and a control policy π is input to the g, and a correction action a′ replacing the action a is output from the g and is actually executed in the environment, wherein a′=g(s, π(s));
• and
• fourthly, the g is trained by means of an evolutionary algorithm or a reinforcement learning method to make sure that data from the unmanned vehicle entity is similar to data from the simulator as far as possible, wherein Σi(si−ui)2 is minimized;
• after the simulator transition correction and transfer, the control policy π obtained through training in the simulator is directly used for the unmanned vehicle entity.