I am developing AI using reinforcement-learning.
It is a game that player should avoid bricks falling from sky.
There are 20 bricks falling to the ground. game screen shot , game play video link
I implemented AI using reinforcement-learning with linear function.
It was hard to choose the best features to get satisfied result.
Anyway it ended up getting best score ever since I tried to make it.
But It's quite weird.
There are some reasons.
- It reached convergence very quickly around 10 training.
- I just used 6 features.
- I couldn't get better score through training more times. (Like I wanted my ai to play like super player. It didn't seem like it though.)
Meanwhile, I tried to use lots of features. I defined features as game screen size(960 * 640).
And I fill those features where brick's relative coordinates from player. I hoped that can make better result than previous one. I trained this method for 7 hours but It didn't work out.
The bottom line is How should I choose the features to make it work best or Is there any different approach to solve it?