Simulated annealing
= Simulated annealing
For a maximization objective, <simulated annealing> accepts every improving proposed neighbor and accepts a worsening move of loss $\Delta$ with <probability> $\exp(-\Delta/T)$ at temperature $T>0$. Decreasing the temperature reduces exploratory moves. Retain the best feasible solution seen. Finite practical cooling schedules carry no general global-optimality guarantee.