I noticed that in the C code distributed with Moses for MERT optimization, there are two routines(/algorithms) for searching the space of possible parameter values, 'optimize_koehn' & 'optimize_powell'.
In Och's 2003 paper on minimum error rate training, he recommends optimizing parameters using Powell's method. However, the code distributed with Moses is hardwired to only call 'optimize_koehn'. Given that you actually have both implementations, but choose to go with optimize_koehn, I assume that Koehn's approach was found to work better in practice. Does anyone recall roughly what the trade offs are between the algorithms? Or, alternatively, are there some drawbacks to Powell's method that make Koehn's a better choice? -Dan _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
