In games like Pinball, the reward system already exists naturally because as points accumulate, the machine can identify what you have to do to get greater numbers: what buttons to press, what movements to do and other keys to have goodresults.However, in games such as "Moctezuma's revenge", the system is not so simple to detect for artificial intelligence.
Everything has to do with the learning process that artificial intelligence systems such as DQN use in this type of games.For example, in video games such as Pinball, DQN made multiple test and error attempts, through which he discarded the possibilities of movement that made him lose points.Something like "If I move to the left I win points, but if I move to the right, no."Then, as attempts progress and the AI memorizes their possibilities, it begins to find the best formulas to succeed.In games such as "Moctezuma's revenge", on the other hand, a simple error in the test causes the character to die, so there is no opportunity to learn the keys based on the repetition and absorption of information, which is the keyof artificial intelligence.
And this is when the participation of babies begins.Deepmind researchers recalled that babies usually look longer at the photographs of images that do not know or have not seen before those who already know, demonstrating that there is something in the simple feeling of novelty that excites babies (and probablyAll humans, if we think about how we make happiness go shopping, among other ways in which money comes to give happiness).
They took advantage of this and added to DQN the ability to get excited and feel attraction for novelty, so every time something new appeared on the screen, this caught their attention and made it acquire the rewards, such as coins and other awards of this typeof games, allowing him to become a star player also of this type of games.