This is our final project submission for the ELENE6885 Reinforcement learnig taught by Prof. Li and Prof. Wang at Columbia University
Text games or popularly known as Interactive Fiction Games have been quite popular for a very long time.
Current reinforcement learning research trends includes exploring and improving these text games which comes with lots of challenges like partial-observability and large action-state space. In this project, we explore the idea of playing text-based games by training a Reinforcement Learning (RL) agent in the textworld environment. We perform our experiments on the game Zork, which is a fantasy based text game. We investigate Deep Reinforcement Learning, with the help of different RL algorithms, such as SARSA and Q-Learning along with different RNN architectures. We also try to draw a comparison between the different agents and their performance. We show that the DDQN agent performs better than an agent that has picked an action randomly.
Additionally, we present a brief finding on different values of
To install the requirements:
pip -r requirements.txt
To run the code, run the following files:
study_compare_diff_rnn.py
study_diff_agents.py
study_effect_of_epsilon.py
study_time_action_space_size.py
TYou can find the project report under:
ELENE6885_2021Fall_textgameZORK_report_rg3332_as6429_sk4824_st3364.pdf