Meta-trained agents implement Bayes-optimal agents
Vlad Mikulik, Grégoire Delétang, Tom McGrath, Tim Genewein, Miljan Martic, Shane Legg, Pedro Ortega
Spotlight presentation: Orals & Spotlights Track 16: Continual/Meta/Misc Learning
on 2020-12-09T07:00:00-08:00 - 2020-12-09T07:10:00-08:00
on 2020-12-09T07:00:00-08:00 - 2020-12-09T07:10:00-08:00
Toggle Abstract Paper (in Proceedings / .pdf)
Abstract: Memory-based meta-learning is a powerful technique to build agents that adapt fast to any task within a target distribution. A previous theoretical study has argued that this remarkable performance is because the meta-training protocol incentivises agents to behave Bayes-optimally. We empirically investigate this claim on a number of prediction and bandit tasks. Inspired by ideas from theoretical computer science, we show that meta-learned and Bayes-optimal agents not only behave alike, but they even share a similar computational structure, in the sense that one agent system can approximately simulate the other. Furthermore, we show that Bayes-optimal agents are fixed points of the meta-learning dynamics. Our results suggest that memory-based meta-learning is a general technique for numerically approximating Bayes-optimal agents; that is, even for task distributions for which we currently don't possess tractable models.