One problem here is that the game cannot test whether you understand the rules, it can only test whether the paths you draw follow the rules. In principle most regular levels (not including the meta-level at the end of each region) could be solved with an extremely tedious brute-force approach of drawing random paths.
So what would it mean for the ML system to figure out the rules? Maybe you could give it the goal to “minimize the number of paths you need to draw in a level to solve it”; but then you separately have to train the system somehow, and what is your training data in a game with handcrafted rather than randomly generated levels?
I’m not sure and that’s the point. I would say your description matches real-life problems quite a bit which makes this applicable to quite a few research topics in AI.
How would you operationalize that?
One problem here is that the game cannot test whether you understand the rules, it can only test whether the paths you draw follow the rules. In principle most regular levels (not including the meta-level at the end of each region) could be solved with an extremely tedious brute-force approach of drawing random paths.
So what would it mean for the ML system to figure out the rules? Maybe you could give it the goal to “minimize the number of paths you need to draw in a level to solve it”; but then you separately have to train the system somehow, and what is your training data in a game with handcrafted rather than randomly generated levels?
I’m not sure and that’s the point. I would say your description matches real-life problems quite a bit which makes this applicable to quite a few research topics in AI.