Skip to content

Scaling All-Goals Updates in Reinforcement Learning Using Convolutional Neural Networks

License

Notifications You must be signed in to change notification settings

fabiopardo/qmap

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Scaling All-Goals Updates in Reinforcement Learning Using Convolutional Neural Networks

Being able to reach any desired location in the environment can be a valuable asset for an agent. Learning a policy to navigate between all pairs of states individually is often not feasible. An all-goals updating algorithm uses each transition to learn Q-values towards all goals simultaneously and off-policy. However the expensive numerous updates in parallel limited the approach to small tabular cases so far. To tackle this problem we propose to use convolutional network architectures to generate Q-values and updates for a large number of goals at once. We demonstrate the accuracy and generalization qualities of the proposed method on randomly generated mazes and Sokoban puzzles. In the case of on-screen goal coordinates the resulting mapping from frames to distance-maps directly informs the agent about which places are reachable and in how many steps. As an example of application we show that replacing the random actions in epsilon-greedy exploration by several actions towards feasible goals generates better exploratory trajectories on Montezuma's Revenge and Super Mario All-Stars games.

The paper can be found on arXiv while videos are available on the website.

Installation

First make sure you have TensorFlow, Baselines, Gym and Gym Retro installed. This code was written for versions 1.11.0, 0.1.5, 0.10.5 and 0.6.0 of these libraries.

To install this package, run:

git clone https://github.com/fabiopardo/qmap.git
cd qmap
pip install -e .

and copy the SuperMarioAllStars-Snes folder to the retro/data/stable directory where Gym Retro is installed.

Usage

First, go to the directory where you wish to save the results, for example:

cd ~/Desktop

By default the training scripts will create a qmap_results folder there.

To train the proposed agent on Super Mario Bros. (All-Stars) level 1.1 you can run:

python -m qmap.train_mario --render

Remove --render to avoid rendering the episodes (videos are saved in the result folder anyway). To train only DQN or Q-map use --no-qmap or --no-dqn. You can also disable both to get a pure random agent.

Similarly, to train the proposed agent on Montezuma's Revenge you can run:

python -m qmap.train_montezuma --render

Or, to learn Q-frames on the proposed grid world use:

python -m qmap.train_gridworld --render

Those scripts produce images, videos and CSV files in the result folder. To plot the values contained in the CSV files, run:

python -m qmap.utils.plot

PDF files are produced which can be kept open and refreshed every 10 seconds using for example:

watch -n10 python -m qmap.utils.plot

To filter which environments or agents to plot, use --witout or --only

To load an agent already trained, run for example:

python -m qmap.train_mario --load qmap_results/ENV/AGENT/RUN/tensorflow/step_STEP.ckpt --level 2.1

where ENV is the environment used to pre-train (for example on level 1.1) and AGENT, RUN and STEP have to be specified.

BibTeX

To cite this repository in publications please use:

@inproceedings{pardo2020scaling,
  title={Scaling All-Goals Updates in Reinforcement Learning Using Convolutional Neural Networks},
  author={Pardo, Fabio and Levdik, Vitaly and Kormushev, Petar},
  booktitle={Thirty-Fourth AAAI Conference on Artificial Intelligence},
  year={2020}
}

About

Scaling All-Goals Updates in Reinforcement Learning Using Convolutional Neural Networks

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages