This paper presents the results of an analysis of accent changes in a corpus of c. 8,000 lexical items. The major results include different analogical paths depending on the word classes (native HL -> LH, Sino-Korean LH -> HL), an Island of Reliability effect (Albright 2002) based on coda consonants, a model of accent change based on weighted constraints employing Jäger (to appear)'s Stochastic Gradient Ascent learning algorithm, and notable deviations from the general LH -> HL trend which take into account sonorant vs. obstruent onsets and token frequencies.