Much additional work could be done in order to highlight more mathematical properties of the graph (and be sure that these properties are real and not aberrations due to the plotting method). This includes points of inflexion, limits, cycles, etc. In addition, when studing a family of curves, it could be nice to emphasize which of these points and properties are constant while the parameter changes.
     In the present version of the Graphing Calculator, as well as in the implementation described in [Fat92], honest plotting is used as an alternative technique to naive or adaptive curve sampling. Unfortunately, honest plotting cannot just replace sampling as too pessimistic bounds sometimes lead to inaccurate results. One way to improve on this would be to systematically display a first version of the graph, then recompute it several times using smaller intervals in order to keep updating the picture while the user is watching it. A more intelligent approach could link honest plotting, adaptive plotting, CR computations, and on-the-fly symbolic computation. The graph could be plotted using essentially CR based adaptive plotting with a simultaneous checking by interval arithmetic (using intervals equivalent to one pixel width). This way, every time honest plotting highlights a possible sampling problem problem, more points could be computed and possibly an alternative method could be tried such as performing a limit calculation or computing a Taylor form inclusion function rather than the natural interval extension [Moo79].
     Possible enhancements of the CR method are manifold: extension to a broader class of mathematical expressions (hyperbolic functions, for example), application in conjunction with adaptive plotting, and adaptation to the manipulation of families of curves and surfaces.
     Considerably more efforts remain to be done towards the implementation of more intelligent plotting packages which would include the many improvements suggested in this paper. In particular, the problems of platform independence and of bi-directional efficient and reliable protocols for numeric and symbolic data exchange have be overcome.