Face images are projected first into a lower dimensional space (S1) using PCA. GAs are then used to evolve the best set of discriminant features by searching all possible rotations of the principal components defining S1 and their corresponding subspaces S2, where dim (S2) <= dim (S1). Evolution is driven by a fitness function defined in terms of performance accuracy and class separation. Face recognition experiments were carried out using 200 FERET test images, whose illumination was different from that of the images used for evolving S2. The results obtained show that the EDF (evolution of discriminant features) performs better than both the the standard eigenface (PCA) and MDF (most discriminant features) methods.