Baby Mario Learns to Jump
Status: Version 1.0
Run via Webstart (b11)
A small experiment in AI Q-Learning. Baby Mario starts on a block, and he needs to practice and learn exactly
how high and far he should jump in order to get to the other block. Uses an adaptive learning AI to teach
Baby Mario - run it 1,000,000 times and see how much he has learned.
|
Features
- Has a working Q-Learner AI that gets better as time passes.
- Can save and load AI files to have already-learned levels of AI.
- Noise can be added to test the capability of the AI - reduce block friction, change wind speed and gravity, etc.
- Has a help menu that says, "Figure it out." Great feature!
- Q-Learner variables (tau, beta, gamma, epsilon) can be changed to explore what's the best method of learning.
- A million frames can be skipped to observe how much he learns after such a time period (he should always be able to land).
Usage
Just try to play around. Baby Mario jumps on his own with no help. See how smart he gets after 1,000,000 iterations. Try randomizing the blocks to see how good he is at landing on absolutely anything after 1,000,000 iterations. Well, some are simply too high for him to reach, but otherwise, he's perfect. If he seems to make a poor choice, he's probably exploring. Try turning changing his explore amount, which is the variable called "epsilon."
Future
Baby Mario has already learned to jump - there's no future for him! However, I continue to use the Q Learner AI method in fun places.