PFSP
Prioritized Fictitious Self-Play (PFSP) is a self-play training mechanism originating from DeepMind's StarCraft II research. It prioritizes matchups against historical agent versions whose win rate against the current agent is close to 50%, thereby avoiding training degradation that would result from repeatedly defeating weaker historical versions.
Timeline (last 90 days)
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.
PFSP preferentially selects opponents against which the current agent's win rate is close to 50%, in order to maximize the learning signal.
Training strategies include curriculum learning (teacher opponents mastering 16 mainstream ladder decks) and PFSP prioritized fictitious self-play league.
All Facts (3)
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.
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%UnverifiedTraining strategies include curriculum learning (teacher opponents mastering 16 mainstream ladder decks) and PFSP prioritized fictitious self-play league.
50%