digger wrote: ↑Tue Oct 08, 2019 11:22 pm

isnt it just an FEA topology optimisation program, the same as what have been around for years? what makes it AI

No. Other optimization routines revolve around either set or statistical algorithms for choices on the optimization variables. The algorithm is largely static (keeps trying the same things over and over), or statistical (simulated annealing, Bayesian optimization, etc)

AI optimization is essentially constantly training a neural network on the system to be optimized at hand, all the while feeding it back results. It works as close to how the human brain does as we have been able to achieve. Not only is it tweaking the allowed variables, it begins to learn how the tweaks will affect the system and where to move next. Additionally, there is usually a "penalty" function introduced that prevents it from hanging in one spot, forcing it to learn more than just a narrow scope of possible solutions. I've seen machine learning optimizations solve 9 degree of freedom design problems and come up with solutions that no human would have ever chosen because the combination defies all convention. But the neural network has "learned" the system from far more dimensions than we have.