I used scene one as a constant variable, so that the good training and the bad training will all be on the same scene. I first tested the good example for the AI to learn and train itself, the result turned out to be surprisingly good; while giving bad examples to the AI worsens the results, so the AI couldn't even finish the rout. Two results showed that bad examples makes it harder for AI to train itself, and this also matches a saying that "garbage in, garbage out". Even though scene two got harder and equally it's harder for me to give good training records and for AI to learn, I still managed to make the AI finished the entire course pretty smoothly.