Single-well application of cyclic gas injection is a promising enhanced oil recovery method especially in fractured and depleted reservoirs. It was previously shown that presence of hydraulic/natural fractures would increase the effectiveness of this method because of the surface area of fractures that supports gas diffusion into the rock matrix. Increased pressure and changes in the physical properties of oil after injection both contribute to improved oil recovery. In this study, a comprehensive analysis of the performance-related design aspects of kcyclic injection of a mixture of N-2, CO2 and CH4 in a hydraulically-fractured well is presented. These three types of gases can contribute to different mechanisms for enhanced recovery because of their unique phase behavior. On the other hand, their costs and availability may be different which may lead to different economic benefits. For the proposed analysis, a large number of injection scenarios were run using a compositional reservoir model that represents the drainage area of a hydraulically fractured well in a low-permeability reservoir. Design parameters such as the injection rate, duration (and volume), soaking duration, economic rate limit, and injected gas composition are varied with uniformly distributed values of each parameter within pre-defined ranges. Both volumetric and economic utilization efficiency factors are analyzed as the performance indicators. Discounted values of incremental oil produced, volume of injected gas, oil price and costs of injected gas for varying project periods of 5, 10, 15 and 20 years are considered for the efficiency calculations. It is observed that the presented method can be a cost-effective enhanced recovery method for hydraulically-fractured wells depending on the proper selection of operational conditions. Injection volume should be kept below 10,000 MCF per cycle. Economic rate limit and soaking period should be optimized. While N-2 is more cost effective than other gases, it mostly provides short-term benefits. The data set is used to train a neural network that can be used to forecast the economic efficiency indicator. Results show that machine learning is an effective approach to develop accurate forecasting models. R-2 values greater than 0.90 are observed for the correlation between real and predicted values for both training and testing, regardless of design considerations of the neural network.