User intent recognition (UIR) enables transfemoral amputees to walk reliably and seamlessly with prosthetic legs. The objective of this paper is to design a UIR system that is optimal in terms of both accuracy and parsimony. We propose the application of two methods to achieve this goal. The first is a filter method, Fisher’s linear discriminant score (FLDS); and the second is a wrapper method, linear discriminant analysis (LDA). Both methods are combined with the evolutionary algorithm biogeography-based optimization (BBO) to find optimal feature subsets. The optimal subsets are then compared with a current state-of-the-art feature selection method, in conjunction with several powerful linear and nonlinear classifiers that are used to identify level ground walking at various speeds. Classification performance is enhanced with a majority voting filter. The best performance is achieved with a multi-class support vector machine that is trained with FLDS/BBO feature subset and that reduces the number of required features by up to about 73% and attains a mean prediction accuracy of 98.94% for amputee subjects. Results show the capability of advanced subset selection methods to construct a UIR system with simultaneous minimum complexity and maximum performance.