Learning Neural Programs To Parse Programs
share
In this work, we study an important problem: learning programs from input-output examples. We propose a novel method to learn a neural program operating a domain-specific non-differentiable machine, and demonstrate that this method can be applied to learn programs that are significantly more complex than the ones synthesized before: programming language parsers from input-output pairs without knowing the underlying grammar. The main challenge is to train the neural program without supervision on execution traces. To tackle it, we propose: (1) LL machines and neural programs operating them to effectively regularize the space of the learned programs; and (2) a two-phase reinforcement learning-based search technique to train the model. Our evaluation demonstrates that our approach can successfully learn to parse programs in both an imperative language and a functional language, and achieve 100 is the first successful demonstration of applying reinforcement learning to train a neural program operating a non-differentiable machine that can fully generalize to test sets on a non-trivial task.
READ FULL TEXT