New research from MIT indicates that AI can be trained to successfully collaborate with other types of artificial intelligence

MIT Technology Review Italy

With the development of self-driving car systems, high-precision surgery and the avant-garde military sector, the collaboration between AI and humans has made great strides. But before this team intelligence can take off, researchers must overcome a serious problem, namely human distrust of AI systems.

Now, at MIT Lincoln Laboratory they have discovered a way to improve AI performance: foster collaboration between completely different AI in a board game called Hanabi. In the wake of independent works by Facebook and those on Google’s DeepMind that affirm the greater validity of training systems that bring together man-AI, the MIT study reaffirms the positive results of cooperative AI, as argued by Ross Allen, researcher of the ‘Artificial Intelligence Technology Group of Lincoln Laboratory in a recent paper presented at the International Conference on Autonomous Agents and Multi-Agent Systems.

To develop cooperative AI, many researchers use Hanabi as a test bed. In this game, you know each other’s cards and not your own and you can only exchange limited information. In a previous experiment, Lincoln Laboratory researchers played one of the world’s best performing AI models for Hanabi with people.. The latter did not appreciate it and called the AI ​​model a confused and unpredictable teammate.

The team then wondered if cooperative AI should be trained differently. Normally, the type of artificial intelligence used, called reinforcement learning, learns how to be successful in complex tasks by discovering which actions yield the highest reward. In this way artificial intelligence players were born unmatched in competitive games such as Go and StarCraft.

But to make it collaborate with other intelligences, AI must not limit itself to maximizing the reward, but must understand and adapt to the strengths and preferences of those who are different. And how does diversity-oriented AI train? The researchers added a new training system, called Any-Play, which included another goal, not new to artificial intelligence, but never extended to collaborative games so far.

In addition to maximizing the game score, the AI ​​must correctly identify the play style of its training partner. To do this, she is obliged to observe the differences in her partner’s behavior. Achieving this goal at the same time requires your partner to learn distinct and recognizable behaviors to convey these differences to the receiving AI.

In the latest version of the research, to assess whether Any-Play training had improved the quality of collaboration, Researchers have expanded participation in the game to more than 100 other AI’s that have been trained by separate algorithms in millions of two-player games.

The pairings coached by Any-Play surpassed all other teams, even when the latter were composed of algorithmically dissimilar partners. Researchers view this type of assessment, so-called inter-algorithmic cross-play, as the best predictor of how cooperative AI would behave in the real world with humans..

“We are interested in understanding the level of interaction when a partner comes into play out of the blue, without any prior knowledge of how they will play. We think this assessment is more realistic when considering cooperation between different AIs, if it can’t be done with humans, ”says Allen.

In their opinion, even if this work didn’t test Any-Play with humans, the system would work. Whether the cross-play scores between algorithms are indeed good indicators of human preference is still a hypothesis. However, to bring a human perspective back into the process, the researchers want to try to correlate a person’s feelings of trust or distrust towards AI to specific targets to be employed in the AI ​​training phase. Finding out about their relationship could help accelerate progress in the field.

Image: Pixabay, Geralt


The post A dialectic of artificial intelligences is possible, indeed necessary first appeared on Technology Review Italia.

#dialectic #artificial #intelligences #Technology #Review #Italia

Leave a Reply

Your email address will not be published.