For my project in our deep learning course, I thought for a long time which deep learning task I should do. We have been doing image classification, natural language processing, and style transfer a lot in class, but I wanted to do something different. I like video games and I like to have fun in what I do. Then in our last session:
So I trained six AIs to do the Lunar Lander task using Deep Q-Networks (DQN) and custom wrappers to change the environment.
How well did the six AIs perform?
After training, I let each model play the game 1000 times.
Notably, model 6 had the highest number of successful landings, with a total of 833. It was followed by model 3 which only had 598 successes.
Although model 6 takes almost twice the time it takes model 3 to play one trial, model 6 had a much higher success rate. Lastly, model 6 also had the highest mean reward, which showcases that its performance is better than the others.