The thing about neural networks is they can generalize from one domain to another. We don' have a million different algorithms, one for recognizing cars, and an

The thing about neural networks is they can generalize from one domain to another. We don' have a million different algorithms, one for recognizing cars, and another for recognizing dogs, etc. They learn features that both have in common. >The "x", the training examples, is what tells the computer what to learn, therefore, what to do once it's trained. If you change the x, the learner can do a different y. Therefore, you're telling the computer what to do. But with RL, a computer can discover it's own training examples from experience. They don't need to be given to it. >I'm still not clear what you're saying; game-playing AI is not an instance of general AI. But it is! The distinction between the real world and a game is arbitrary. If an algorithm can learn to play a random video game, you can just as easily plug it into a robot and let it play "real life". The world is more complicated, of course, but not qualitatively different.