- direct iterative methods:
-
generalized iterative scaling
-
guess values of the Lagrange multipliers
-
for each constraint j, keeping lambdai constant,
estimate the j'th multiplier
-
if all constraints satisfied sufficiently, stop else repeat
previous step
-
the above procedure is guaranteed to converge for equality
constraints (Darroch and Ratcliff)
- via Maximum Likelihood:
-
use ME framework to get the form of the probability distribution
-
estimate parameters of the distribution using ML
-
nice mix of two paradigms where models are not randomly chosen
and estimation is done using available data