RLAssist is available under the Apache License, version 2.0. Please see the LICENSE file for details.
Rahul Gupta, Aditya Kanade, Shirish Shevade. "Deep Reinforcement Learning for Syntactic Error Repair in Student Programs", AAAI 2019.
RLAssist uses the dataset developed in our previous work: "DeepFix: Fixing common C language errors by deep learning". The dataset was collected from an introductory programming course at Indian Institute of Technology, Kanpur, India using a programming tutoring system called Prutor. It is available under the Apache 2.0 license, courtesy Prof. Amey Karkare and his research group. If you it for your research, kindly give due credits to both Prutor and DeepFix.
If you are using Ubuntu 16.04.5 LTS (not tested on other distributions) and have conda installed, you can run source init.sh which creates a new virtual environment called rlassist and installs all the dependencies in it.
Furthermore, it downloads, extracts, and preprocesses the student submission data for you into the required directory structure.
Otherwise, run source init.sh and manually fix any failing steps.
To start training on one of the folds with default hyper-parameters, you could run the following.
python -O neural_net/agent.py data/network_inputs/RLAssist-seed-1189/bin_0/ data/checkpoints/RLAssist-seed-1189/bin_0/
To run a configuration without expert demonstrations, use -ge 0 -sr flags.
You can also control the number of worker threads using -w flag.
Please see the details of other flags in neural_net/agent.py.
Training script will periodically write to a log file in the logs directory.
To use the trained model to fix programs in the test dataset, take the value of the checkpoint from the log file (say 23495767) and run the following.
python -O neural_net/agent.py data/network_inputs/RLAssist-seed-1189/bin_0/ data/checkpoints/RLAssist-seed-1189/bin_0/ -eval 23495767