Fuzzy input and output pairs are presented to fuzzy-weighted neural networks under a supervised training regime. The training employs an initial backpropagation phase followed by a genetic algorithm phase. In an archtypical case, the former is used on peaks of triangular fuzzy inputs and outputs from the training set and the latter is applied to the entire training set. Resulting weights are triangularly shaped fuzzy, crisp numbers or a combination of both. Mappings to be learned are chosen according to several patterns for manipulating the level or degree of fuzziness: leave intact, e.g. in the identity mapping; increase; decrease; or increase over a portion of the domain and decrease over the remainder. Impact on training time, weight types and configurations, response to inputs not in the training set and other properties are alluded to together with comparisons to use of genetic algorithms alone or combined with table look-up. Mappings comparable to some of those studied abstractly are employed in concrete applications characterizing aspects of behavior in the ``RoboKid" problem, in which a child (with or without mild mental retardation), a robot, or ``team robot" performs instruction following tasks involving moving objects about in a laboratory setting.}