w9PcJLyb / LuxAI2021-10th-Place-Solution

10th Place Solution for LuxAI2021 Competition

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

LuxAI2021 10th Place Solution

Lux AI Competition - https://www.kaggle.com/c/lux-ai-2021

The architecture I used was inspired by this post

Major changes:

  • Separate model to predict city actions {BUILD_WORKER, RESEARCH, NOTHING}
  • Transfer actions for units {TRANSFER_NORTH, TRANSFER_SOUTH, TRANSFER_WEST, TRANSFER_EAST}
  • Augmentations (Flip, Rotate)

agents/IL_1689.tar.gz - My best agent, 10th place in the competition

How to train the agent:

  1. download episodes, example: https://www.kaggle.com/code/robga/simulations-episode-scraper-match-downloader
  2. train the unit model: python3 imitation_learning/train_units.py --episode_dir "path_to_episodes"
  3. train the city model: python3 imitation_learning/train_cities.py --episode_dir "path_to_episodes"
  4. copy agents/IL_1689.tar.gz and update the models there.

About

10th Place Solution for LuxAI2021 Competition

License:Apache License 2.0


Languages

Language:Python 100.0%