Comparative Analysis of Reinforcement Learning Algorithms on TORCS Environment

In this study, reinforcement learning algorithms are compared in TORCS simulation environment. In this simulation environment, the goal is to finish the track as soon as possible by controlling the car. The agent decides actions by using highlevel observations from the environment. For this goal, two reinforcement learning algorithms (Deep Deterministic Policy Gradient (DDPG) and Deep Q Network (DQN)) are used and the results are compared and analyzed. Since the action space is continuous, DDPG algorithm performed better as expected. However, we were able to show that DQN algorithm also gives comparable results.