Skip to content

Latest commit

 

History

History
32 lines (17 loc) · 3.36 KB

README.md

File metadata and controls

32 lines (17 loc) · 3.36 KB
Raw

Deep Quality Value (DQV) Learning

DQV-Learning: a novel faster and stable synchronous Deep Reinforcement Learning algorithm.

This repo contains the official code and models of the paper Deep Quality-Value (DQV) Learning which has been presented at the NeurIPS Deep Reinforcement Learning Workshop, Montreal (CA), 2018: https://arxiv.org/abs/1810.00368

Given a standard RL setting, unlike Temporal-Difference (TD) Reinforcement Learning algorithms such as DQN and DDQN, DQV aims to learn directly a Value function with the TD(λ) update rule:

which estimates can then be used to learn the state-action pairs of the Q-function via:

When approximating the two update rules with a parametric function approximator like a neural network, both functions can be expressed as regression problems that can be minimized through gradient descent by optimizing

and

To check the benefits of these update rules that are able to learn up to 3 times faster and better than DQN and DDQN as presented in the paper be sure to:

  • Install the requirements list in your favourite virtualenv with pip install -r requirements.txt

  • The directory ./models/ contains the weights of the Value and Quality networks which have obtained the results that are reported in the paper for the game Pong. If you aim to test one of these pre-trained networks you can run the ./models/DQV_Trained.py script. Be sure to use the weights that match the appropriate Open-AI environment.

  • If you aim to train a model from scratch, or adapt DQV on a different DRL problem, you can find the code for training the agents in ./train/. The .train_job.sh script allows you to choose which Open-AI environment you want to train your agents on, and which exploration strategy to follow. So far the code only supports e-greedy and Maxwell-Boltzman exploration. The script will call DQV_learning.py which is the code that has been used to train the networks that are present in ./models/, and that matches with what is reported in the paper. Please note that if you aim to train an agent on a different game than Pong some modifications to the code might be needed.

Training on the game Pong lasts ≈ 24 hours on a GTX 1080 GPU machine, the OpenAI agent should be defeated after ≈ 400 episodes (≈ 8 hours of training), whereas the environment should be fully solved in ≈ 600 episodes, with DQV obtaining a reward of ≈ 21.