We propose a neural network method for the generation of symbolic expressions using reinforcement learning. Usually, the symbolic form expressed in terms of a calculus (propositional, first-order, lambda, etc.) is deemed comprehensible by humans and it is necessary as far as the acceptance of neural networks is concerned.According to the proposed method, a human decides on the kind and number of primitive functions which, with the appropriate composition (in the mathematical sense), can represent a mapping between two domains. The appropriate composition is achieved by an agent which tries many compositions and receives a reward depending on the quality of the composed function. Naturally, the learning agent (which in our case is a recurrent neural net) must perform the credit assignment task. Results are encouraging concerning the derivation of simple arithmetic expressions.