[Cover] [Contents] [Index] |
Page 265
Figure 6.20 The probability for class replacement. As T is decreasing, only small increases of energy can be accepted.
figuration w is visited infinitely often, the distribution of generated configurations is guaranteed to converge to the Gibbs distribution. The second theorem concerns the temperature T of SA. It states that if the decreasing sequence of temperatures satisfies the following condition:
(6.50) |
where
(6.51) |
and m is the number of lattices (i.e. pixels), then convergence can be guaranteed. However, to decrease temperature T in terms of Equation (6.50) is too slow for practical applications. One may apply another temperature cooling function f defined as (Dubes and Jain, 1989):
(6.52) |
The initial value of temperature T0 is usually set to a value of 2 or 3.
The parameter Ninner in Figure 6.19 requires further discussion. Generally, the value of Ninner can be set as small as 10m or as high as 200m, depending on how many configurations are intended to be tried at each temperature.
[Cover] [Contents] [Index] |