This study discusses cognitive process in solving constraint satisfaction problem in terms of both operation and information acquisition strategies by examining human process data with eye mark tracking solving Sudoku. Generally a Sudoku problem is defined as a CSP where a cell corresponds to a variable, and its domain of values is a set of the numbers. One of the major inference strategies to solve it is constraint propagation based on local consistency that resolves conflicts within a local variable set, and a problem that is solvable only by resolving the two-variable consistency is classified as "easy". However, the result of human data does not support this naive problem definition and the strategy. It can be explained by another problem representation where a number corresponds to a variable, and a cell to a domain value. We check the cognitive validity of this representation and the strategies based on it.