What if one day, AI could evolve like living beings?

Progress in artificial intelligence is continuing. In fact, American researchers are at the origin of a small revolution. Recently, their work has led to the development of an AI capable of transmitting "useful modifications" to future generations.

A new machine learning technique

For several years, artificial intelligence has impressed humanity. Today, there are RNs capable of diagnosing diseases, identifying asteroids that are potentially dangerous to Earth, or predicting earthquakes, among other things. We usually talk about the evolution of AI, but in reality, the same evolution is about the scientific work, not the AI itself.

Indeed, artificial intelligence has not yet managed to mimic the evolution of life. None of them has thus transformed itself into a response to its environment, by learning from its mistakes or by passing on mutations to their descendants. Typically, scientists develop specific problem specific AI's in a separate way. Nevertheless, a recent study published in nature Communications on October 6, 2021 embodies a small revolution in AI. The team at Stanford University (USA) who initiated the work explains that they have developed an AI capable of imitating life (see video at the end of this article).

In their publication, the scientists refer to a new machine learning technique. This allows AI to transform over time, depending on the complexity of its environment. In this way, the researchers in the study hope to elucidate some of the principles governing the relationships between environmental complexity, advanced morphology, and the ability to learn intelligent control.

A morphology passed on to future generations

Several learning agents compose and define the morphology of AI in order to adapt and find solutions. The AI is then placed in a virtual environment and must move independently. After several trials, the agents will determine which mutations may be useful for, for example, travelling long distances. These mutations will then be passed on to subsequent generations, who will benefit from the same morphology as their eldest. On the other hand, this will not be the case for apprenticeships.

By generating several morphologies to solve just as many problems, the researchers have gone against what RNs usually do. They test a very wide range of solutions and choose the most effective. In addition, the new AI increases its level of intelligence in line with the complexity of the environment. According to scientists, environments that require escaping, manipulating objects or exploring provide opportunities for AI to learn more quickly.

For project leaders, this work could support the robotics industry. More experiments should be undertaken to better understand how learning and evolution together create sophisticated relationships between the complexity of the environment, morphological intelligence, and the ability to learn about control tasks.?