The computer to which the laziness. As the AI learned how to escape from the manmade task

Компьютер, которому лень. Как ИИ научился отвлекаться от поставленных человеком задач

“Rise of the machines” is still a fantastic story, very far from reality, which, however, does not prevent to prosper in other subjects. Here and now

Great Dumatel! We have come so far. Have you calculated the ultimate question of life, the Universe, and everything?

— Oh, no. I was watching TV…

In the early 1990s, this dialogue from the works of the English science fiction writer Douglas Adams seemed very funny. The world’s most powerful artificial intelligence, more than ten billion years, hangs in front of the TV instead of to solve the main task set before him the Universe. Well, if not absurdity!

Procrastination was considered a purely human weakness, when it was attributed to the computer, that was funny. Today, scientists involved in the development of artificial intelligence over such issues and I don’t laugh.

Joke of thirty years ago has become a real problem, which is not so easy to solve. American non-profit research structure Open AI tested, in which the artificial intelligence agent had to go through a virtual maze to take an object, which he pointed how important and valuable. The experiment is stalled at the stage when the maze appeared on the TV, and the remote agent.

Instead of moving toward the goal, look for the designated object, the agent just stuck in front of the TV, endlessly switching channels. In another test he had a virtual laser as a shooter game, and instead of looking for the object, he was just shooting at walls to see where you’re going.

When you imagine the possible consequences of such a failure on the stage of the practical application of artificial intelligence systems, on the back runs a chill. What will end it, if during napaloni operations agent, who must find and recognize the point of impact in the human body, suddenly “distracted”. Or, for example, the agent providing auto-negotiation of the vehicle, do something “more interesting” while driving at full speed.

In fact, these risks is due to the fact that for learning artificial intelligence use of the effect is akin to a human surprise

As strange as it may sound, wanting to force the AI to learn and explore new spaces and territories, scientists taught him to be surprised. Based on the previous observations, it predicts that there should be “around the corner”, no matter in the virtual maze, on the road or in the coronary artery of the operated patient if the forecast is justified, there is something else, there is pure novelty.

The system of “reinforcement learning”, about which there is a speech, is designed so that by performing the right action, the agent receives conditional promotion. But if he’s looking for something or paves the route, then the correct action is to move on to new spaces, and not to stagnate or laps. Therefore, the encouragement he receives when “surprised”.

The trouble is that the artificial intelligence is able to find other sources of surprise, not related initially its task. And then there is the paradoxical situation where the system does absolutely not what she wants.

A group of researchers from Google Brain, Deep Mind and the Higher technical school of Zurich seems to be suggested a method of learning in which the agent will not procrastinate as lead they will have no desire to “wonder”, and the effect of occasional curiosity. However, technopessimism already talked about the fact that the person everything becomes more difficult to control samooborony artificial intelligence.

The more complex and more functional the system becomes, the less chance that the test will identify all possible failures. Not to mention the fact that, together with properties similar to curiosity, the AI becomes self-motivated in principle, not dependent on the will guide person. Don’t overreact — “rise of the machines” is still a fantastic story, very far from reality.

However, sometimes science fiction writers predict the future. Something of what in the late 1980s — early 1990s, wrote Douglas Adams, it seems, has come true.

Share Button