3 ms·
AI Designs Quantum Physics Experiments Beyond What Any Human Has Conceived
- gus_massa 5y agoResearch paper: https://arxiv.org/pdf/1509.02749.pdf https://arxiv.org/pdf/1509.02749.pdf The algorithm is not very clear. Is the code available? It looks like a mix of randomized brute force, with some heuristic to keep some "interesting" examples. From the article: > S7) Learning algorithm > In the second example involving cyclic operations, the algorithm extends its own set of basic elements autonomously, based on the properties of the longest cycle. It saves elements that have large cycles and experiments with non-trivial coupling between different degrees-of-freedom. Additionally, elements already learned can be forgotten to improve variability and prevent dead-ends, as some of them might even have negative effects on the probability of finding new experiments. The decision of which elements are forgotten at which times is purposefully random. Even though it would be possible to weight the elements for past usefulness, it would introduce a bias on similar solutions that we wanted to prevent.
- martincmartin 5y agoIs it a genetic algorithm?
- gus_massa 5y agoIt's very difficult to be sure, but it looks like a genetic algorithm without crossover???