Meta-trained agents implement Bayes-optimal agents

TL;DR

This paper empirically demonstrates that memory-based meta-learning agents behave similarly to Bayes-optimal agents and share computational structures, suggesting meta-learning can approximate optimal Bayesian inference across various tasks.

Contribution

The study provides empirical evidence that meta-trained agents implement Bayes-optimal strategies and share similar computational structures, supporting the idea that meta-learning approximates Bayesian inference.

Findings

Meta-trained agents behave like Bayes-optimal agents.

Meta-trained and Bayes-optimal agents share similar computational structures.

Bayes-optimal agents are fixed points of meta-learning dynamics.

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 might serve as a general technique for numerically approximating Bayes-optimal agents - that is, even for task distributions for which we currently don't possess tractable models.