Myoelectric signals are a standard input for volitional control of prosthetic devices. As an information-rich signal, feature selection plays a decisive role in the performance of motion classification. In this paper, we evaluate feature selection in the classification of simultaneous motions produced from combinations of wrist and elbow flexion/extension, radio-ulnar pronation/supination, and hand opening/closing aiming to determine a common set of recommendations for the implementation of motion classification from EMG signals for prosthetic control. Chow-Liu trees and forward feature selection are used as the methods for selecting features, and six different classification algorithms are evaluated as the wrapping component. We analyzed the performance of different linear and non-linear kernel algorithms in terms of the accuracy with respect to the feature selection, observing that feature selection was critical for improving accuracy levels to above 95%. Chow-Liu trees demonstrated to be a strategy that enables a combination of a low number of iterations with comparable accuracy to what is achieved with a forward selection search. In addition, we found a trend for waveform length and entropy as the most relevant types of features to consider and found evidence that simultaneous motion classification should be handled using non-linear classification approaches. Our study serves to improve the feature and algorithm selection for surface EMG signals in the classification of simultaneous motions generating a viable approach in the recognition of combinations of actions from the upper limb.