Google AI taught to be "extremely aggressive" in stressful situations

Last year, the famous theoretical physicist Stephen Hawking said that the development of artificial intelligence would be "either the best or worst event for all mankind." We all watched "the Terminator" and we all imagine how apocalyptic hell may be our existence, if such a self-aware AI system like Skynet, one day decides that humanity no longer needs it. And the latest results of the new AI system from the company DeepMind (owned by Google) only once again remind us of the need to be extremely careful when manufacturing robots of the future.

In studies conducted at the end of last year, AI from DeepMind acquired and demonstrated, regardless of what is laid to his memory, and defeated the best players in the world . In addition, he perfected his skills and mimicry of the human voice.

In the latest test, the researchers checked "her desire" to cooperate. Tests showed when the AI DeepMind "feels" that lose, for the avoidance of loss he begins to choose a new and «extremely aggressive» strategy. The Google team spent with AI, 40 million sessions in a simple computer game Gathering, where the player was required to collect as much desired fruit. AI DeepMind operated with two "players agents" (a blue cube and a red cube). Google engineers were given the task to organize a competition between the «agents» and collect as much virtual apples (green squares) as possible.

From time To time, while "agents" could easily assemble fruit, available in large quantities, everything went smoothly. But as soon as the supply of apples decreased, the behavior of "agents" became "aggressive." They began to actively use the tool (laser beam), which helped to dislodge the enemy behind the game screen and then to collect all the apples

Interestingly, knocking out of the enemy from the field with a laser beam no further awards are proposed. Broken the enemy is just some time is outside of the screen, while more successful, the enemy was able to gather more virtual apples.

If the "agents" did not use laser beams, then theoretically the number of their apples would be the same. It's basically happened when as "agents" was used more low-level and "less intelligent" options DeepMind. More aggressive behavior, sabotage and greed began to show only when Google began to use more and more complex forms of DeepMind.

When the researchers used as "agents" more than a simple network DeepMind, between them noted the presence of "atmosphere more friendly coexistence on the playing field." However, when the managing agents was transferred more and more complex forms of networking, AI became more aggressive and began to try in advance to knock the opponent from the playing field to first faster to get to the lion's share of production from virtual apples. Google scientists assume that the more intelligent an agent, the more effective it is able to learn, to adapt to the environment and the methods available, and eventually come to use the most aggressive tactics to win.

"This model shows that the result of learning and adaptation to environmental conditions is a manifestation of certain aspects of human behavior" — says Joel Z. Leibo, one of the researchers who conducted this experiment.

"Less aggressive behavior occurred only when training and staying in a relatively safe environment, with less chance of consequences after certain actions. Greed, in turn, was reflected in the desire to overtake and to collect all the apples."

After the "harvest" DeepMind was offered to play another game called Wolfpack. This time it was attended by three AI-agent: the two played the role of wolves, and the other the role of prey. Unlike the game Gathering, new game, strongly contributed to the cooperation between wolves. First, it is easier to catch prey, and secondly, if both "wolf" were near pounded the prey, they both receive a certain reward, regardless of who it caught.

"the Idea is that mining can be dangerous. In addition, can lone wolf and able to drive, but there is a risk of loss in an attack by scavengers", — explains the command.

"But if both of them chase the prey together, then they are better able to protect her from the scavengers and thus generate a higher reward."

In General, from game Gathering AI DeepMind learned that aggression and selfishness are the most effective strategies for getting desired result in a particular environment. Wolfpack of the same system realized that cooperation, as opposed to the individual attempts could be the key to a more valuable reward in certain situations. Although described above test environment are only basic computer games – the main message is clear now. Take a different AI with competing interests in real-life situations, place in one environment, and maybe if their tasks are not kompensiruet common goal, the result may be a real war. Especially if people like one of elements in the achievement of this goal will be eliminated.

As an example, just imagine the traffic lights, AI and unmanned vehicles, attempt to find the quickest route. Everyone performs their tasks with the goal of getting the safest and most effective results for companies.

Despite the "infant" years DeepMind and the lack of any third-party critical analysis of its capabilities, the results of its tests suggest the following thoughts: even if we create, it does not mean that robots and AI systems will automatically put the desire to put our human needs above their own. We have to "plant kindness" in the nature of machines and to anticipate any "loopholes" that may allow them to reach those laser beams.

One of the main theses of the initiative group OpenAI, to study the ethics of artificial intelligence as something sounded in 2015 as follows:

"Today's AI systems possess awesome, but highly specialized capabilities. And most likely, we still long time will keep in your improvement. At least as long as their ability to solve virtually every intellectual challenges will not be greater than the human. It is difficult to imagine what level of benefit you will be able to bring AI to human-level companies, as it is difficult to imagine what damage it can cause to society while the neglect of its creation and use."