The process of reducing the number of attributes from an information system is known as attribute reduction problem. The action of attribute reduction problem is an important step of pre-processing in data mining. In the attribute reduction process, reduction is performed with considerations for minimum information loss. Having a dataset with discrete attribute values, it is possible to find a minimal subset of the original attribute set in rough set theory. Great Deluge algorithm optimizes this problem by controlling the search space using lower boundary “Level”. This paper presents a modification of Great Deluge algorithm for rough set attribute reduction wherein the “Level” is nonlinear. In the modified model, the “Level” is increased by a value that is calculated based on the quality of the current solution for each iteration. An alternative neighborhood structure assists the nonlinear Great Deluge to improve the quality of the method. The standard datasets available in a UCI machine-learning repository are employed to examine the proposed method. The accuracies of the classifications are investigated using ROSETTA. The performance of the proposed method is evaluated by comparing the results of nonlinear with linear Great Deluge algorithm and other available approaches in the literature. Statistical tests are performed to analyze the results. Experimental results show promising results of the proposed method compared to other available approaches in the literature.