search

QiXuanWang / LearningFromTheBest

This project is to list the best books, courses, tutorial, methods on learning certain knowledge
8 stars 1 forks source link

AlphaStar: Grandmaster level in StarCraft II using multi-agent reinforcement learning #4

Open QiXuanWang opened 4 years ago

QiXuanWang commented 4 years ago

Source: https://deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-StarCraft-II-using-multi-agent-reinforcement-learning

Ref2: https://aisc.ai.science/events/2019-12-09

Problems: StarCraft is very complicated game

Innovations:

  1. Supervised (imitation) learning for agent initialization
  2. Population Based Multi-agent reinforcement learning “League” with main and exploiter agents
  3. Combines a lot of different technology: Transformer, LSTM core, Pointer Network etc.

Comments: Population Based Multi-agent RL has been a very hot topic since then. Unlike AlphaGo/AlphaZero, MCTS is not used?

QiXuanWang commented 4 years ago

Ref: (very little information) https://paperswithcode.com/paper/alphastar-an-evolutionary-computation

Github that implemented PBT: https://github.com/MattKleinsmith/pbt