varibad-zintgraf-2019-variational-autoencoder
IN premise — summaries/2026/08/24/wiki-Meta-learning_computer_science.md
Created 2026-08-25T02:58:49+00:00
VariBAD (Zintgraf et al., 2019) is a model-based meta-reinforcement learning method that uses a variational autoencoder to capture task information in internal memory and conditions the policy on that task.
Summary
VariBAD is an approach where a reinforcement-learning agent learns to compress "what task am I on?" into a small internal memory and then uses that memory to decide how to act. In this system it sits as a foundational definition, so any later claim comparing methods, explaining task generalization, or attributing results to VariBAD is measured against this VAE-based memory-and-conditioning mechanism as the core design.