Method and system for optimizing reinforcement-learning-based autonomous driving according to user preferences

A computer system for executing a robot autonomous driving learning method based on deep reinforcement learning, comprising:

• at least one processor configured to execute computer-readable instructions included in a memory,
• wherein the at least one processor executes the steps comprising:
• applying a plurality of different autonomous driving parameters simultaneously to a plurality of robot agents in a simulation by randomly assigning predetermined setting values of the autonomous driving parameters to the plurality of robot agents by a system or a manager, to enable the plurality of robot agents to perform autonomous driving that fits the assigned autonomous driving parameters without retraining;
• acquiring, in real time, a sensor value relating to a movement of a robot from a robot agent of the plurality of robot agents applied with the different autonomous driving parameters;
• inputting the sensor value and an autonomous driving parameter of the different autonomous driving parameters assigned to the robot agent to a neural network comprising a plurality of neural network parameters for learning an adaptable robot autonomous driving policy to perform autonomous driving, where the learned adaptable robot autonomous driving policy is adaptable to various autonomous driving parameters without retraining;
• updating the plurality of neural network parameters to train the robot agent on the adaptable robot autonomous driving policy;
• and
• after said updating the plurality of neural network parameters, optimizing the autonomous driving parameters by applying a feedback from a human viewing two driving images of a robot to which the autonomous driving parameters are set differently for each of the two driving images respectively, wherein the robot in said each of the two driving images is the trained robot agent,
• wherein the optimizing of the autonomous driving parameters comprises:
• modeling a preference for the autonomous driving parameters using a Bayesian neural network, and
• generating a query for obtaining the feedback from the human by selecting the autonomous driving parameters corresponding to the two driving images based on an uncertainty estimate of the modeled preference to minimize a number of queries.