AlphaStar
An AI system developed by DeepMind to play StarCraft II, employing multi-agent self-play and league training mechanisms. In 2019, it reached the level of professional players, representing a significant breakthrough in applying deep reinforcement learning to complex real-time strategy games.
Timeline (last 90 days)
For small-scale projects, fully reproducing AlphaStar-level training is unrealistic, but "periodically training an exploiter targeting the current policy and adding it to the league" is a simplified version that can be implemented on single-machine resources.
PFSP preferentially selects opponents against which the current agent's win rate is close to 50%, in order to maximize the learning signal.
The AlphaStar project was designed by the DeepMind team for StarCraft II, and its league-style self-play framework maintains an opponent pool containing the main agent, historical versions, and exploiters, and uses prioritized fictitious self-play (PFSP) to sample opponents.
The success of AlphaGo Zero and AlphaStar has made the self-play paradigm highly praised, but successful cases often rely on enormous computing resources and carefully designed training.
PPO is one of the underlying algorithmic choices for large-scale game AI projects such as OpenAI Five (Dota 2 AI) and AlphaStar.
DeepMind developed AlphaGo and AlphaStar (StarCraft), while OpenAI developed a Dota 2 team battle AI
DeepMind released AlphaStar in 2019, and its league consisted of three types of agents: main agents, main exploiters, and league exploiters.
The concept of league play was applied at scale in DeepMind's StarCraft II AI, AlphaStar.
League play uses population diversity to avoid the rock-paper-scissors problem of cyclic dominance.
DeepMind's AlphaStar used StarCraft as a landmark environment for testing AI's comprehensive capabilities, and listed coping with the Fog of War as one of its core technical challenges.
5 more timeline events
All Facts (16)
AlphaStar在2019年在《星际争霸II》欧服达到大师级水平（前0.15%），通过模仿学习与多智能体强化学习的结合实现
90%UnverifiedFor small-scale projects, fully reproducing AlphaStar-level training is unrealistic, but "periodically training an exploiter targeting the current policy and adding it to the league" is a simplified version that can be implemented on single-machine resources.
50%UnverifiedThe success of AlphaGo Zero and AlphaStar has made the self-play paradigm highly praised, but successful cases often rely on enormous computing resources and carefully designed training.
50%UnverifiedThe AlphaStar project was designed by the DeepMind team for StarCraft II, and its league-style self-play framework maintains an opponent pool containing the main agent, historical versions, and exploiters, and uses prioritized fictitious self-play (PFSP) to sample opponents.
50%UnverifiedPFSP preferentially selects opponents against which the current agent's win rate is close to 50%, in order to maximize the learning signal.
50%UnverifiedPPO is one of the underlying algorithmic choices for large-scale game AI projects such as OpenAI Five (Dota 2 AI) and AlphaStar.
50%UnverifiedDeepMind developed AlphaGo and AlphaStar (StarCraft), while OpenAI developed a Dota 2 team battle AI
50%UnverifiedThe concept of league play was applied at scale in DeepMind's StarCraft II AI, AlphaStar.
50%UnverifiedDeepMind released AlphaStar in 2019, and its league consisted of three types of agents: main agents, main exploiters, and league exploiters.
50%UnverifiedLeague play uses population diversity to avoid the rock-paper-scissors problem of cyclic dominance.
50%UnverifiedDeepMind's AlphaStar used StarCraft as a landmark environment for testing AI's comprehensive capabilities, and listed coping with the Fog of War as one of its core technical challenges.
50%UnverifiedOriol Vinyals is the VP of Research at DeepMind and was a central figure in projects such as AlphaStar and AlphaCode.
50%UnverifiedThe breakthrough in Atari deep reinforcement learning laid the methodological foundation for subsequent projects such as AlphaGo and AlphaStar.
50%UnverifiedDeepMind's AlphaStar, OpenAI Five, and other milestone projects are built on top of games.
50%UnverifiedAlphaStar reached the Grandmaster level in StarCraft II in 2019.
50%UnverifiedMilestones in modern game AI research include DeepMind's DQN demonstrating deep reinforcement learning on Atari games in 2013, AlphaGo defeating the Go world champion in 2016, OpenAI Five defeating professional teams in Dota 2 in 2019, and AlphaStar reaching Grandmaster level in StarCraft II in 2019.
50%