It is usually assumed that the choice behavior of bees for floral colors is influenced by innate preferences only for the first flower visits prior to any experience. After visits to rewarding flowers bees learn to associate their colors with a reward. This learning process leads to an acquired preference for the trained colors that has been believed to dominate over previous experiences and over innate preferences. This work investigates how bumble bees (Bombus terrestris) chose among artificial flowers of different colors after they had been extensively trained to other colors. The bees chose novel colors according to their similarity to the trained color if the trained color was similar to some of the test colors. This was true also if trained colors and test colors were well distinguished, so their color choice reflected generalization between colors. If the test colors were so different from the trained color that no generalization took place, choice behavior was not affected by the trained color and reflected innate preferences. The differences in choice frequencies could not be explained by physical properties of the test colors other than the dominant wavelength, a parameter taken to reflect hue perception. Preferred dominant wavelengths correspond to those observed in naive bumble bees and honeybees. Thus bumble bees show innate preferences for certain colors not only prior to color learning but also after intensive learning when choosing among very different novel colors. Color choice among similar colors, however, is controlled by generalization from the learned color.