VIME: Variational Information Maximizing Exploration
What this work does
Adds an approximate KL change in a Bayesian dynamics model to the external reward, using variational neural-network inference and a practical second-order approximation.
Connection to discovery
An implementation baseline for training evidence-seeking behavior from belief updates rather than raw environmental surprise.
Evidence
Sparse-reward continuous-control experiments compare several underlying policy optimizers and report external returns separately from intrinsic rewards.
Scope and limitations
Variational approximation and model misspecification affect the bonus. Better control return is not direct evidence of scientific mechanism recovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.