In English

Path Planning using Reinforcement Learning and Objective Data

Tian Xia ; Zijian Han
Göteborg : Chalmers tekniska högskola, 2018. 74 s.
[Examensarbete på avancerad nivå]

With the rapid development of autonomous driving vehicles, decision making for path planning has become advanced and challenging topics. Traditional planning and control methods are usually limited by the difficulty to find good solutions, so deep machine learning has become engineers’ focus in order to solve these problems. Several related works using reinforcement learning have been done in the simulation environment TORCS. This thesis will focus on training an vehicle to learn driving at certain target speed on high way condition without collision. A complete learning structure is designed for vehicle system, and a hierarchical learning algorithm will be used with deep reinforcement learning methods. Deep Q learning is used to learn option level of policy, and deep deterministic policy gradient is used to learn primitive action level of policies. Neural networks are used to approximate the value functions. The training results are tested on various set up of opponents vehicles on the track, with the probability of damage recorded and compared.

Nyckelord: reinforcement learning, TORCS, policy, option, DQN, DDPG.



Publikationen registrerades 2019-01-10. Den ändrades senast 2019-01-10

CPL ID: 256451

Detta är en tjänst från Chalmers bibliotek