The Reinforcement Learning Environment

Policy type instruction [DEMO]

---

Software Architecture

During Training

During Testing

---

How to run the code?

0. Git clone the project https://github.com/tirthankar95/NumberLearningInChildren_RL_NLP.git and go inside one directory. Check if you have permission to clone the repository.
1. Download glove embeddings wget http://nlp.stanford.edu/data/glove.6B.zip and unzip it unzip glove*.zip.
2. Pull the environment(docker image) for running the code from docker-hub using. sudo docker pull tirthankar95/rl-nlp:latest.
3. Go inside the container sudo docker run -it -v $(pwd):/NumberLearningInChildren_ML tirthankar95/rl-nlp /bin/bash
4. Run from the main directory ~ python3 -W ignore RL_Algorithm/PPO/ppo.py &> Results/console.log &. This is used to train the model.
5. Run from the main directory ~ python3 -W ignore RL_Algorithm/PPO/ppo_post_run.py &. This is used to generate results after the model has been trained.

Analysis

1. You can generate your own code for custom analysis. To reproduce our results copy Results/final_score_* and Results/train_model_* to Analysis folder.

2. Append {a} to train_model_*.json{a} in Analysis folder.

   Where,
   {a} -> {"", "seed_0", "seed_1", ... } depending on whether it's the {1st, 2nd, 3rd, ...} run of the same configuration.
   

3. Run python3 Plot_redx.py inside the container.

4. To create consolidated graphs modify train_config.json and test_config.json, by trying out all the combinations below.
   model $\epsilon$ {0, 1, 2, 3}
   instr_type $\epsilon$ {0, 1}
   And then do step 1 & 2, after dumping all the necessary files in the Analysis folder do step 3. This will give consolidated graphs as depicted in the paper.

Copyright (c) 2024 Tirthankar Mittra