In this project, LunarLander-v2 (LL) was chosen as the reinforcement learning problem to be solved. The lander initially spawns at the top of the screen and begins to fall with gravity, until it either crashes or comes to a halt. The agent’s aim is to select actions in such a way that maximises the total amount of reward it receives within the duration of the episode. To complete the task however, the agent must be allowed to learn how to control the lander, so that it consistently finishes on the landing pad without crashing. The task is considered complete when the agent obtains a mean reward of 200, over 100 consecutive episodes.
Every time step the agent receives an observation from the environment containing eight values, each corresponding to one of the items in the list below.
The agent must then choose one of the four possible discrete actions listed below.
The agent receives a reward of +10 every time one of the lander’s legs touches the ground, as well as a negative reward for moving away from the landing pad. The agent receives a reward of −0.03 for firing the side engine and −0.3 for firing the main engine. The terminal states are those in which the landing craft has either crashed (due to landing upside down or too quickly) or has landed successfully, resulting in rewards of −100 and +100 respectively. Finishing on the landing pad carries an additional bonus reward of between +100 and +140.
