Pre-training Framework for Improving Learning Speed of Reinforcement Learning based Autonomous Vehicles

摘要

Reinforcement learning based autonomous vehicles have the disadvantage of long learning time. The paper proposes a pre-training framework for improving learning speed of autonomous vehicles (PRELSA) in reinforcement learning. PRELSA framework pre-learns the agent's neural network before actual learning by pre-initializing the agent's policy gradient neural network. Simulation results show that PRELSA framework improves learning speed by about 20 percent compared to existing learning method.