Pommerman

Folder Structure

• data_gen : Code related to data generation for training
• training : Code related to train the neural network
• docker : Docker for running neural networks
• web_app : for human interaction

To Do List

NIPS Competition

• The env is PommeTeamCompetition-v0 There will only be collapsing walls if we cannot easily determine a winner from normal play.

State

• 11x11 board:
  • hardwall, box, passage, player, partner, enemy, bomb, flames
  • fog
  • bomblife, bomb strength

Questions:

• How is partial information encoded?
• How is Power-Ups encoded? Bomb, Range, Can-Kick
• How many past information should I store?

Data Augumentation:

• flip the board across all the dimensions
• simulate the partial visibility??
• switch the agents, enimies
• Simulate the end cases by randomly generation of all possible end conditions?

Other codebases

• dataset : https://github.com/tambetm/pommerman-baselines
• C baseline : https://github.com/m2q/nips2018-agent

References:

• https://towardsdatascience.com/introduction-to-various-reinforcement-learning-algorithms-i-q-learning-sarsa-dqn-ddpg-72a5e0cb6287
• https://towardsdatascience.com/introduction-to-various-reinforcement-learning-algorithms-part-ii-trpo-ppo-87f2c5919bb9
• TRPO: https://arxiv.org/pdf/1502.05477.pdf
• PPOA: https://arxiv.org/pdf/1707.06347.pdf

Other

• https://help.github.com/articles/basic-writing-and-formatting-syntax/