https://github.com/miffyli/minecraft-bc

Submission code of UEFDRL team to NeurIPS 2019 MineRL challenge (5th place)
https://github.com/miffyli/minecraft-bc

imitation-learning keras machine-learning minecraft

Last synced: about 1 month ago
JSON representation

Submission code of UEFDRL team to NeurIPS 2019 MineRL challenge (5th place)

• Host: GitHub
• URL: https://github.com/miffyli/minecraft-bc
• Owner: Miffyli
• License: mit
• Created: 2020-04-27T13:38:36.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-11-13T17:46:09.000Z (about 4 years ago)
• Last Synced: 2023-03-06T13:22:21.915Z (almost 2 years ago)
• Topics: imitation-learning, keras, machine-learning, minecraft
• Language: Python
• Size: 21.5 KB
• Stars: 11
• Watchers: 4
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Playing Minecraft with behavioural cloning

This repository contains the final ranked submission of UEFDRL team to the [MineRL 2019 challenge](https://www.aicrowd.com/challenges/neurips-2019-minerl-competition),

reaching fifth place.

Long story short: Behavioural cloning on the provided dataset, _i.e._ predict what actions humans would take. No RNNs.

See this paper for full details: [Playing Minecraft with Behavioural Cloning](https://arxiv.org/abs/2005.03374).

## Contents

Code is in the submission format, and can be ran with the instructions at [submission template repository](https://github.com/minerllabs/competition_submission_starter_template).

`requirements.txt` contains Python modules required to run the code, and `apt.txt` includes any Debian packages required (used by the Docker image in AICrowd evaluation server).

The core of our submission resides in `train_keras_imitation.py`, which contains the main training loop. 

## Running

[Download](http://minerl.io/dataset/) and place MineRL dataset under `./data`. Alternatively point environment variable `MINERL_DATA_ROOT` to the downloaded dataset.

Run `train.py` to train the model. Afterwards run `test.py` to run the evaluation used in the AICrowd platform. This code prints out per-episode rewards.

After 200 games, the average episodic reward should be around 10-13. The results very from run-to-run, and we also

noticed our local evaluations having consistently lower score than on AICrowd platform (achieved +15 results).