Pequi - Reinforcement Learning Competition - CEIA 2021 (1st Place)

Members:

Competition Envoriment: https://github.com/bryanoliveira/soccer-twos-env
Starter Code: https://github.com/dlb-rl/rl-tournament-starter
Solution Presentation (Portuguese): https://youtu.be/8C91NqZV4yk?t=7456
Solution Presentation Slides: https://docs.google.com/presentation/d/1X9TG1GYlHuVzvuVGsUv8bPfMYdgxGImLq0JfLakHxwc/edit?usp=sharing

Solutions:

Goiaba: (0.47 Win-Rate vs Baseline)
GoiabaV2: (0.6 Win-Rate vs Baseline)
Goiabão: (0.81 Win-Rate vs Baseline)

Requirements

• Python 3.8
• See requirements.txt

Usage

• Clone this repository
• Install the requirements: pip install -r requirements.txt
• Run python example_random.py to watch a random agent play the game
• Run python example_ray_team_vs_random.py to train team vs team against a random agent using Ray RLLib

You may also run this environment on Colab.

Tournament submission

To submit an agent for the competition you must follow this instructions:

• Implement a class that inherits from soccer_twos.AgentInterface and implements an act method
• Fill in your agent's information in the README.md file (agent name, authors & emails, and description)
• Test your agent module as described in the next section
• Compress your agent's module folder as .zip and e-mail it to bryanlmoliveira@gmail.com.

See example_player_agent/ or example_team_agent/ module for reference.

Testing/Evaluating

Use the environment's rollout tool to test your module before submission:

python -m soccer_twos.watch -m example_player_agent

You may also run your agent against our pre-trained baseline (download). Extract the ceia_baseline_agent folder to this project's folder and run:

python -m soccer_twos.watch -m1 example_player_agent -m2 ceia_baseline_agent