-When "increasing" a value, multiply it by two until the expected effect is observed. After the first time the value becomes too large (i.e. the behavior is unstable or the mechanism overshoots), reduce the value to halfway between the first too-large value encountered and the previous value tested before that. Continue iterating this "split-half" procedure to zero in on the optimal value (if the response undershoots, pick the halfway point between the new value and the last value immediately above it - if it overshoots, pick the halfway point between the new value and the last value immediately below it). This is called an term:`exponential search`, and is a very efficient way to find positive values of unknown scale.
0 commit comments