This paper describes a well-motivated and elegant method for optimizing execution of neural network computation graphs using reinforcement learning. The paper is well-written, however there is no open source implementation, hence results may be difficult to reproduce.