I'm a huge fan of video games and Mario. I've been learning decision trees / machine learning and this is quite a fascinating video that connects the two by using machine learning to complete a Mario level.
From what I understand, this scripts tests and tries and then remembers each set of screen states and the correct responses. Therefore, it won't be as good at generalizing in the sense that if you loaded another level, it won't perform as well. It will need to learn from trial and error what the correct move is and then remember then repeat.
One of the limitation of this method is that once it has learned that something works well, it is not likely to unlearn it even if it might not be the best move. You can insert some sort of 'mutation' that has a random chance of doing something slightly different but it might not be more efficient overall and might take longer to learn and complete the level.
At the end of the day, it depends on what your objective is to do for machine learning. I'm only beginning to dabble in it and it is a complex but very interesting area.
Source code can be found here: http://pastebin.com/ZZmSNaHX
Credit to SethBling and you can connect with him with the following:
Login to post comments