Mengchen originated from the Temple of Concave Fei
Qubits | Official account QbitAI
An AI tool that has made physicists ecstatic, open sourced on GitHub!
It is called Φ-SO, which can find hidden laws directly from the data, and directly give the corresponding formula in one step.
The whole process does not require supercomputing, and a notebook can solve Einstein's mass-energy equation in about 4 hours.
The results came from the University of Strasbourg in Germany and the Australian Commonwealth Scientific and Industrial Research Organization Data61 Department, according to the first paper, the research took 1.5 years and received extensive attention from the academic community.
Once the code is open sourced, the star is also rising rapidly.
In addition to physicists calling Amazing out, researchers from other disciplines rushed to explore whether the same method could be transferred to their field.
Reinforcement learning + physical constraints
The technique behind Φ-SO is called "deep symbolic regression" and is implemented using recurrent neural networks (RNNs) + reinforcement learning.
First input the previous symbol and context information to the RNN, predict the probability distribution of the latter symbol, repeat this step, you can generate a large number of expressions.
At the same time, physical conditions are incorporated into the learning process as prior knowledge to avoid AI from coming up with formulas with no practical meaning, which can greatly reduce the search space.
Then introduce reinforcement learning, so that the AI learns to generate formulas that best fit the original data.
Unlike reinforcement learning, which is used to play chess, manipulate robots, etc., symbolic regression tasks only need to care about how to find the best formula, not about the average performance of the neural network.
So the rules of reinforcement learning are designed to reward only the first 5% of candidate formulas, and find the other 95% without penalty, encouraging the model to fully explore the search space.
The research team used classical formulas such as the damping harmonic oscillator analytical expression, Einstein's energy formula, and Newton's universal gravitational formula to do experiments.
Φ-SO can restore these formulas from data 100%, and the above methods are indispensable.
Compared with other methods into MLP, Φ-SO also performs better outside the training range.
In the end, the research team said that although the algorithm itself still has some room for improvement, their first task has been changed to using new tools to discover unknown physical laws.
GitHub:
https://github.com/WassimTenachi/PhySO
Thesis:
https://arxiv.org/abs/2303.03192
Reference Links:
[1]https://twitter.com/astro_wassim/status/1633645134934949888
Reference link: [1] https://twitter.com/astro_wassim/status/1633645134934949888
— End —
Qubits QbitAI · Headline number signed
Follow us and be the first to know the latest scientific and technological trends