Expression Inference is a parsimonious, comprehensible alternative to semi-parametric and non-parametric classification techniques such as neural networks, which generates compact symbolic mathematical expressions for classification or regression. This paper introduces a general framework for inferring symbolic classifiers, using the Genetic Programming paradigm with non-linear optimisation of embedded coefficients. An error propagation algorithm is introduced to support the optimisation. A multiobjective variant of Genetic Programming provides a range of models trading off parsimony and classification performance, the latter measured by ROC curve analysis. The technique is shown to develop extremely concise and effective models on a sample real-world problem domain.