Loading
Warm-up Round: Completed Round 1: Completed #reinforcement_learning #classroom
11.6k
138
4
255

๐Ÿ“š  RL research papers and resources to help you get started
๐Ÿ‘พ  Info on challenge rewards and Steam keys

๐Ÿ•ต๏ธโ€โ™‚๏ธ Introduction

Dr. Derk's Mutant Battleground is our newest multi-agent RL challenge build around the Dr Derk's Gym game. It is a MOBA-style RL environment for python that runs on the GPU. In this challenge, you can benchmark against other players online.

Multi-online battle arena games have become increasingly popular. Dota 2, League of Legends and Vainglory are favourite among the online gaming community. Advanced RL research even saw five neural networks defeat amateur players in Dota 2. The benefit of training and creating multi-agent RL models are enormous!

Dr Derk provides a unique "lite" environment as it runs entirely on the GPU. You can run hundreds of โ€œarenasโ€ simultaneously. In an hour, you can get about 23,000 games on a single machine. The game employs โ€œrealโ€ artificial intelligence. It also employs both discrete and continuous action space. You can build and train your Derkling as you like and defeat the opposition. What makes this challenge interesting is the tournament-style showdown where you can visually see your RL-model fight another and see your Derkling win! 

Want to know how to make your first submission in under 15 minutes, read on! โฌ‡๏ธ

๐Ÿ”— Gym installation instructions and documentation here!

๐Ÿ‘พ What is this challenge?

Once upon a time, a scientist named Dr Derk engineered a creature to have a neural network brain. This creature, called Derkling, is powered by a recurrent neural network. It has 60 senses and 12 actions. Your job is to train them and win!

You get to control a team of three "Derklings" to fight an opponents team. For each battle, the Derklings are randomly equipped with three items each. They have 60 senses, 5 actions and 22 tweakable rewards. You get 4 points per opponent Derkling you kill, and 13 points to kill their Statue; the winner is the player with the most points. Let the best Derkling win!

๐Ÿ’ช Getting started

You can create your first submission in under 15 minutes by following these steps:

1. Click here to fork the starter kit. 

2. Create tag push to your repository with the prefix submission-

3. Now, you can see the details of your submission in the repository's issue tracker.

Click here to see more details on making submissions. Happy gaming!

๐ŸŽฎ Reward

Participants making their first successful submission will receive a Steam key to the complete Dr Derk's Gym game. For more information, read our post.

โš– Evaluation metric

The bot submitted by the participant will be evaluated against a combination of easy, medium and difficult bots. The difference in score between the participant bot and the evaluator bot will together decide the leaderboard score. 

The leaderboard score will be calculated as a weighted sum across the different bots: 

  • The easy bot gets a weight of 0.1
  • The medium bot gets a weight of 0.3
  • The hard bot gets a weight of 0.6

The formula used for the calculation is:

\(x = {-b \pm \sqrt{b^2-4ac} \over 2a}\)

๐Ÿ—‚ Competition Structure

  • Warm-Up Round
    The warm-up round evaluates submissions on an easy medium and difficult bot. The scoring metric will be the weighted average of the scores of the games against bots of different complexity.
  • Round One
    Round 1 will witness the same scoring metric as the warmup round, consisting of easy, medium and difficult bot. 
  • Round Two
    Round 2 will be a tournament styled round. It will consist of three parts, an instantaneous evaluation where the bot will be evaluated on the easy, medium and hard bots used in the warmup round and the first round. The second part will be a weekly tournament where the top 20% of the bots will be paired up in a bracket tournament based on their rating. The third and final part will be a final tournament where we will choose the best 20% of bots throughout the round and have them face off against each other.

๐Ÿ“… Timeline

  • Warm-Up Round: Dec 21st - Jan 15th
  • Round 1: Jan 15th - Feb 15th 
  • Round 2: TBD

๐Ÿ† Prizes

We do not yet have a prize pool and are actively searching for sponsors. If you are interested in sponsoring this competition with prizes for the winners please reach out by sending a mail to connect@aicrowd.com! 

๐Ÿ‘ฅ Team

The organizing team consists of:

Participants

Getting Started