Experimental designs were compared using stacked-layer feed-forward neural networks. Several traditional three-level designs and uniform designs were investigated using three-factor linear and nonlinear models. The prediction error was found to be inversely proportional to the number of experiments. Uniform designs displayed better performance than traditional three-level designs for the same number of experiments. The sum of squares of prediction errors was generally smaller for uniform designs. The performance difference between three-level designs and uniform designs was attributed to the number of factor levels. This was confirmed by further investigation on random designs with more factor levels.