darkmatter2222 / Halite-RL-DQN

Deep Rnforcement Larning to play Halite

Halite-RL-DQN 🤖

Exploring Deep Rnforcement Larning to play Halite, specifically the version of Halite on:
https://www.kaggle.com/c/halite
https://github.com/Kaggle/kaggle-environments/tree/master/kaggle_environments/envs/halite

Baby Steps First 👶

This was my first attempt at reinforcement learning, before chomping off a huge bite, lets train a simple bot. Yes, I studied Q Learning and Q tables, and jumped right into Deep Reinforcement Learning.:laughing:
https://github.com/darkmatter2222/Halite-RL-DQN/blob/master/susman_rl/dqn_bots/find_the_dot_v0/dqn.py
This is a POC of sample code from the tf_agents documentation and my idea for a good challenge similar to Halite::thinking:
https://www.tensorflow.org/agents/overview

(Speed reduced for GIF representation)
https://www.youtube.com/watch?v=-33ehDNT3kY

Goal: The white dot to find the green dot
Avoid: Falling off the map or taking too many steps

• On a 5x5 Grid and a typical 2018-2020 CPU/GPU, training time can take ~4k steps and 10-15 minutes. w/ >95% Win rate.
• On a 6x6 Grid and a typical 2018-2020 CPU/GPU, training time can take ~12k steps and 40-60 minutes w/ >95% Win rate.
• ...
• On a 15x15 Grid and a typical 2018-2020 CPU/GPU, training time can take ~10M-20M steps and 1-2 days w/ >95% Win rate.:woozy_face:
  • Can anyone help improve this?

Try 'Baby Steps' yourself? 🖥️

1. Config directories here:
   1. https://github.com/darkmatter2222/Halite-RL-DQN/blob/master/susman_rl/dqn_bots/find_the_dot_v0/config.json
2. Set board size here:
   1. https://github.com/darkmatter2222/Halite-RL-DQN/blob/master/susman_rl/environments/find_the_dot_v0/env.py
3. Train bot here:
   1. https://github.com/darkmatter2222/Halite-RL-DQN/blob/master/susman_rl/dqn_bots/find_the_dot_v0/train_model.py
4. Evaluate bot here:
   1. https://github.com/darkmatter2222/Halite-RL-DQN/blob/master/susman_rl/dqn_bots/find_the_dot_v0/execute_trained_model.py