Simulated Annealing: a stochastic optimization method inspired by nature. As "temperature" decreases, random fluctuations (that allow it to escape local minima/maxima) decrease and the particle resorts to hillclimbing. No gradient calculation required. #RStats #rayshader

(See if you can spot the two particles going "We're coming! Wait for us!!")

@Flexi23 This animation does not show the full power of SA, because not many of them have reached the global maximum. However, it seems this is because of non-optimally chosen parameters (not enough time spent at very high temperature).

@certhionyx good ol simulated annealing

@wgeary Picking out the ridge lines is a nice emergent property too. Sweet.

I never got simulated annealing but it’s amazing how much it clicks with the animation. It’s fantastic!

@DFintelligence tu cherchais une visualisation en stream hier pour expliquer l'optimisation de fonctions objectives

@saforem2 How do they know to climb up if no gradient is being computed? Or do you mean the global maximum?

@saforem2 As the "temperature" lowers, the probability of a random movement to a lower potential being accepted decreases. As T -> 0 this probability also goes to zero, and the algorithm effectively reverts to pure hillclimbing.

The only thing that is missing is a voice over for the last two data points climbing up the hill 😂 Great work!