Learning to Reason Mathematically (original) (raw)

We introduce the simplification of mathematical expressions as a sequential task whose solution requires understanding the structure of the expressions. We do not assume any expert information and develop a curriculum learning algorithm that makes learning in a space with a highly sparse reward signal possible. Graph Neural Network is used to represent the expressions and we show via an intermediate task that it has sufficient expressive power to keep the necessary information for the simplification. The proposed algorithm is able to learn the simplifying sequence of actions from scratch by solving a curriculum of expressions with increasing complexity.