Experiments in this study show two ways in which children's use of inflected forms in Korean deviate from adult usage, and provides a unified account. Either by avoiding suffixation or by applying minimal modification when suffixation is required, younger children aged 4;2 to 5;8 produce forms that are faithful to a base form of noun paradigms. I argue that this is due to grammatical preference among younger children that requires non-alternating paradigm. Specifically, the constraints requiring outputs faithful to a base form, BD-IDENT constraints, are very highly ranked in child grammar, thus they avoid alternations in early stage of learning. In the later stage of learning, aged 6;2 to 7;9, older children allow alternations reflecting suffix-specific preference, but not lexical-specific preference of obstruent-final nouns' variants among adults' production. As for adults, the current study using picture description task also found that they show lexical-specific and suffix-specific preference for the alternation of obstruent-final nouns, as reported in previous corpus study (Choi 2004) and nonce word experiment (Jun and Lee 2007). Assuming highly ranked BD-IDENT constraints (McCarthy 1999), learning simulation in this study investigated whether the attested learning stage can naturally emerge in the way of re-ranking sets of constraints. I fed the model with the frequency of alternations from an adult speech corpus and checked if the model can re-weight constraints just by exposing the rankings to probabilistic data from the corpus. The model was able to predict the attested three learning stages, in that predicted outputs match with the attested production data in each age group in the experiments. Therefore, the current study suggest that the three stages of learning Korean noun paradigms are predicted by training a grammar with statistical distribution of alternations.