Simulated Annealing vs. Genetic Algorithm
When I read about simulated annealing, I considered implementing it instead of a genetic algorithm. But I decided not to. Why? Simulated annealing relies on the assumption that the best layout is next to the second best layout, which is next to the third best layout, etc. But this is likely to be untrue. The second best keyboard probably looks nothing like the best keyboard. But genetic algorithms avoid this problem, through repetition. The “all star” round contains many very good layouts, so it is more likely to converge on the best layout.
But when I saw some results, I had to reconsider. Simulated annealing is seemingly much faster. But how can we get it to converge on the best possible layout? Could we do something like simulated annealing, but then repeat it and pool all the best layouts and evolve those using a genetic algorithm?
I ran some empirical tests, and it turns out that simulated annealing is indeed many times faster than a comparable genetic algorithm. Chris Johnson sent me some source code, and it turns out that it is indeed much faster.
Simulated annealing works sort of like a “smart” genetic algorithm with a pool size of only one. Instead of just making any old mutation, it looks for good mutations to make. This allows much faster convergence. But this strategy, as well as the genetic algorithm strategy, can sometimes skip over very good layouts, or even the best. I will explain in a later post, coming soon.