Quiz dossier / Reinforcement Learning

AlphaStar Deep Dive: The Nature Paper (Vinyals et al., 2019)

Question-by-question drill on the AlphaStar paper: the two-phase training pipeline, network architecture, RL update (TD(lambda), V-trace, UPGO, KL), league training with PFSP and exploiters, fairness constraints, and transfer to complex multi-agent systems.

22 questionsEstimated 45 min0 flagged

Session status

Draft

Answered: 0/22
Score: 0/22

Question 1

1 point(s)

AlphaStar's training pipeline has two main phases. What are they, in order?

Supervised (imitation) learning from human replays, then multi-agent reinforcement learning in a league

Reinforcement learning from scratch, then supervised fine-tuning on human replays

Unsupervised pretraining, then evolutionary search only

Hand-coded scripting, then behavioural cloning