**5. Conclusions**

We presented a framework for designing a UIR system. We used experimental data collected from three able-bodied subjects and three above-knee amputee subjects to classify four and three different gait modes, respectively. Overlapped windowing with frame length 250 ms and increment 50 ms provided a good trade-off between classification performance and real-time computation. Several efficient TD and FD features were extracted from data frames to form the feature set. We performed feature selection in two steps. First, we excluded non-informative features with poor classification performance and high computational effort. Second, we used MOO to find an optimal feature subset from the remaining features to obtain a UIR system that was both parsimonious and accurate. For this purpose, GMOFS, a novel embedded multi-objective feature selection algorithm, was proposed and compared with four evolutionary MOOs on the basis of normalized hypervolume and relative coverage. Classification results confirmed the competitive performance of GMOFS. Several classifiers were trained with the optimal feature subsets that were selected by MOO, and SVM-RBF and MLP were found to be the best classifiers for UIR. The outputs of the classifiers were input to an MVF to improve classification accuracy and chattering between the identified classes.

For future work, more above-knee amputee subjects will be involved in data collection and classification. In addition, we will include other daily-life activities such as incline walking, stair ascent and descent, standing and sitting, etc. It is also of great interest to consider other informative features for classification, such as wavelet transform coefficients. Finally, it would be of interest to compare GMOFS with other state-of-the-art MOO methods, and to apply GMOFS to other MOO problems.

**Author Contributions:** Investigation, G.K.; Methodology, G.K. and H.M.; Project Administration, D.S.; Resources, H.M.; Software, G.K.; Writing—Original Draft Preparation, G.K. and H.M.; Writing—Review and Editing, D.S.; Funding Acquisition, D.S.

**Funding:** This research was supported by National Science Foundation grants 1344954 and 1536035, and a Cleveland State University Graduate Student Research Award.

**Acknowledgments:** The authors would like to thank Elizabeth C. Hardin from Cleveland Veterans Affairs Medical Center who contributed to the experiments of collecting data.

**Conflicts of Interest:** The authors declare no conflict of interest.
