The performance optimization of many man-made systems belong to simulation-based constrained optimization (SBCO), where the evaluation of both the performance and the constraint have no closed form expression and are based on simulation. The simulation-based estimate of both the performance and the feasibility are usually time-consuming and noisy. So it is of great practical interest to study how to efficiently find a good and feasible design. Constrained ordinal optimization (COO) is one of the few methods that solve SBCO. We follow the idea of COO to develop a crude model for performance evaluation and a feasibility model to estimate the feasibility. The existing selection rules in COO only search within designs that are observed as feasible. We first show that this could lead to an efficiency lower than blind picking in some cases. Then we develop an improved COO (ICOO) that searches within the entire design space and compares designs efficiently. Both theoretical and numerical results demonstrate that ICOO provides an efficient way to solve SBCO and beats the existing selection rules in COO, including blink picking with feasibility model and horse racing with feasibility model. We hope this study sheds insight to solving simulation-based constrained optimization in general.