The authors studied an actor-critic algorithm, which learns both how to act and how to estimate the value of its actions, in Starpilot and Craftax.
Small training runs predict reinforcement learning gains, but local reversals remain
Tests in Starpilot and Craftax predict performance and training choices at up to 100 times more compute, using a single-agent Elo metric.
Big Tech
Zeyu Zheng · Clare Lyle · Khimya Khetarpal · Will Dabney
Google DeepMind
Research Digest··2 min read
Zheng and colleagues test whether small reinforcement learning experiments can predict the results of much larger training runs.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.