AI is not just a synthesis of what humanity already knows, but also a search for something new. One of the major breakthroughs in AI was the victory of a computer over a human in chess without any training from a human. The computer simply played against itself and focused on the final result (win or lose) and adapted its strategy accordingly. This type of learning is called Reinforcement Learning - in my opinion, the most interesting area in AI.
Unfortunately, practical application is only possible where experiments can be conducted relatively endlessly, over and over again. As you can understand, this is difficult to implement in self-driving cars - we are not ready to crash expensive equipment repeatedly, but in simulators, it is quite feasible. That’s why more and more various emulators for robots (synthesized virtual worlds with virtual objects) and cars (creating non-existent roads) are appearing. By the way, this is why Tesla has a significant advantage - it has a vast amount of data collected for free by car owners.
But today, that’s not what I wanted to talk about. I wanted to mention that reinforcement learning is also applied in physics. For example, the Φ-SO project is an attempt to find physical laws from scratch using data. It’s an incredibly interesting approach where the machine generates hypotheses, checks them against reality (whether the formula is correct or not), and adjusts its strategy accordingly.)
Paper: https://arxiv.org/abs/2303.03192
Code: https://github.com/WassimTenachi/PhySO
Thread: https://twitter.com/astro_wassim/status/1633645134934949888