Transformer Based Reinforcement Learning for Games
This repository contains experimental models, written in PyTorch, which incorporate transformers in the Deep Q-Learning tasks, to see if they perform better than the RNN based version (DRQN) or simple DQN.
Requirements
* OpenAI Gym
* PyTorch >= 1.0.0
* Python 3.6+
* Conda (suggested for building environment etc)
* tensorboardx==1.9
* tensorflow==1.14.0 (non-gpu version will do, only needed for tensorboard)
(environment.yml provides detailed list of dependency)
How to run experiments?
Currently, we experiment with the cartpole environment, and experiment with
the three different algorithms, DQN, DRQN (using LSTM) and a transformer based model called DTQN.
The repo is structured in the following mannner
-src/
|-config_*.py (config files of a particular algorithm)
|-model_*.py (model definition for a particular algorithm)
|-train_*.py (training file of a particular algorithm)
|-memory.py (action replay memory buffer)
-out/
|-trace_*.txt (traces obtained by different algorithms)
To run a particular algorithm (say DQN) one can do python train_DQN.py this will generate the trace for that algorithm.
Results
We have performed multiple experiments for each of the algorithms - DQN, DRQN and DTQN. Each algorithm was trained for 5000 episodes and we ran 10 different instances for each of the algorithms with random initialization.
The folowing pictures illustrate the plots of scores over episodes for different runs.
