As a result of trying to play the game "Montezuma's Revenge" whose main character dies immediately with Google's artificial intelligence "DQN"

Google's artificial intelligence development company "DeepMind" has defeated the Go champion "AlphaGo"At the same time, artificial intelligence that learns the game himself and can improve to more than humans"DQN"We are also developing. Once you start the program, DQN that you can practice the game like "Space Invaders" and "Block Break" without permission, even if you do not modify them, you can get to expert level, the game where the hero dies quickly A movie that is fighting hard at "Montezuma's Revenge" is released.

Playing Montezuma's Revenge with Intrinsic Motivation - YouTube


The play screen of "Montezuma's Revenge" looks something like this. There is a hero wearing red clothes and a hat in the room, and a skull is rolling on the floor.


The main character jumping and jumping from 3rd floor to 2nd floor. What the characteristic of this hero is ...


To die soon. If you are in Mario you jump and landing is a fun thing, basically jumping from a high place from Montezuma's Revenge's hero will die.


And even if you hit a skull rolling on the floor ... ...


The body disappears and dies. Because it dies soon, it is Montezuma's Revenge that DQN struggled the most among the 2,600 Atari game.


So what should I do? I mean ....


A hero who gets downstairs using a ladder if you do.


Jump ......


I will hold on to Paul.


After jumping and landing on the right end of the second floor, go down the ladder ......


Move to the left end of the room.


Get the key.


You can move to another room. Although it is the hero who walks quickly ... ...


It died after hitting the wall represented by the dotted line.


Level 1 has 24 rooms in all, and it is a mechanism to get the key and move from room to room.


Among them, DQN said that they trained 100 million times room 1 and room 2.


At this time, DQN performed different movements each time. In other words, we explored ways to clear the room with various patterns.


In the second, purple room ......


Get a sword.


There were scenes like the following in the purple room.


The hero who jumps with Pyeongchin and moves on in the room in a shaky way.


However, he died immediately after hitting the wall.


Try again ... ....


death.


Furthermore, I will pull it and change the movement.


I wonder if I can go right in the moment the wall disappears and I can successfully escape ......


Somehow back.


A wall appeared again and died.


Finally through the wall ......


I was able to get off at the ladder.


......, and so far it seemed to be in a state of training, the result after having trained 100 million times from here. I will approach the skull rolling on the floor.


Peeon suddenly jumped over.


Then get the key.


I tried to go to the next room with Ya ....


It is a death from a jump → landing on the floor.


As I managed to clear the first room, I used the ladder to get down.


And to the purple room which was a difficult point.


Although it is good to have started walking at the moment the walls disappear ......


For some reason turning direction. The wall will reappear at that moment.


Of course I will die.


If a human plays, it is not such a difficult scene, but I repeat dead.


Successfully finished through the wall.


And to the floor further down in the ladder.


Going to the right ......


Another room with another atmosphere appeared.


Get the key using Paul.


Go along Paul ......


Go down further in ladder without becoming a flow of jump → landing → death.


It was DQN who was repeating the movement of "What is you doing?" At all times, dead and reviving, but it seems that he finally succeeded in clearing the two rooms.


In addition, it is said that 15 rooms have already been searched out of 24 rooms at level 1 at the moment.


· Continued
The boundary line between Google's artificial intelligence "DQN" that can play better than humans and games that can not be done - GIGAZINE