In this paper, an algorithm is developed to deal with the discrete optimization problem. The algorithm approximates any design continuous domain with finite number of discrete points and employs single and multi-level search to reach near-to-global optimum. In single level search, one orthogonal array is used to model any given search domain. In multi-level search, two or more orthogonal arrays are coupled in series and used to model the search domain. The number of design levels are increased with the number of arrays via different coefficients. The tolerance synthesis problem with optimum process combination is revisited to compare our method with well-established algorithms such as Simulated Annealing (SA) and Sequential Quadratic Programming (SQP). The effect of algorithm parameters: different structure combinations, reducing move factors, weighing factors and column assignments on optimum for single and multi-level search are investigated. Results indicate the capability of the approach to reach near-to-global optimum in about 5.20%–19.5% of time taken by other methods. A survey of available algorithms is given and the method is validated with few test cases.