In the previous recipe, Partitioning Monte Carlo simulations for better pmap performance, we more or less guessed what will make a good partition size. We tried a few different values and saw what gives us the best results. However, it's still largely guesswork since just making the partitions larger or smaller doesn't give consistently better or worse results.

This is the type of task that computers are good at. Namely, searching a complex space to find the function parameters that result in an optimal output value. For this recipe, we'll use a fairly simple optimization algorithm called simulated annealing. Similar to many optimization algorithms, this is based on a natural process: the way molecules settle into low-energy configurations as the temperature drops to freezing. This is what allows water to form efficient crystal lattices as it freezes.

In simulated annealing, we feed a state to a cost function. At each point, we evaluate...