We introduce a general framework for the definition of function classes. Our model, which is based on nondeterministic polynomial-time Turing transducers, allows uniform characterizations of FP, FP NP , FP NP [0(logn)], FP$ P , counting classes (#-P, #-NP, #-coNP, GapP, GapP NP ), optimization classes (max-P, min-P, max-NP, min-NP), promise classes (NPSV, #f ew -P, c#-P), multivalued classes (FewFP, NPMV), and many more. Each such class is defined in our model by a scheme how to evaluate computation trees of nondeterministic machines. We study a reducibility notion between such evaluation schemes, which leads to a necessary and sufficient criterion for relativizable inclusion between function classes. As it turns out, this criterion is easily applicable and we get as a consequence, e.g., that there is an oracle A, such that min-P^ % #-NP^ (note that no structural consequences are known to follow from the corresponding positive inclusion). Int. J. Found. Comput. Sci. 2000.11:525-551. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 08/23/15. For personal use only.