SIMULATED-ANNEALING allows the agent to take steps with decreases in the objective function with some probability.  The probability depends on the schedule AND on the degree of “badness” of the move.  In this sense all bad moves are not created equal.  As more steps are taken in the algorithm, the probability of a backward move decreases and converges to only taking moves with increasing utility.  If the rate of annealing is correctly set, the algorithm is guaranteed to find a solution if one exists.
	Note that it picks a move at random.  The probability transition is only used for moves with negative value.  Any positive valued move (which increases the objective function’s value) is taken with probability 1 if it is chosen, no matter how big or small the gain is.
	To think about: what types of problems would this algorithm work well on?  Work poorly on?