Predicting the Performance of Code using Machine Learning

The goal of this research paper is to see the potential of machine learning to predict the performance of a given program.

The goal of this research paper is to see the potential of machine learning to predict the performance of a given program. Doing this could allow for faster predictions than compiling and properly testing a given program, especially with high compile-time optimization, allowing for faster optimization of a program. To accomplish this, we trained a model consisting of a TreeLSTM and an MLP on a dataset sourced from CodeChef in order to predict the time and memory taken by a given program. After training, we received an average percent error of 22.43% and 1261.07% for time and memory respectively which is fairly poor overall but somewhat fine for the time. As such, this paper concludes that machine learning can potentially be used to accomplish this goal but would likely need a more involved and thought through implementation to work effectively.