How to Stay Curious while Avoiding Noisy TVs using Aleatoric Uncertainty Estimation
What this work does
Predicts both transition means and variances, subtracting a learned aleatoric term from prediction-error curiosity.
Connection to discovery
Useful when designing discovery rewards that should favor learnable uncertainty rather than irreducible noise.
Evidence
Noisy-image transitions, MiniGrid and game environments test action-dependent and naturally occurring stochastic distractions; comparisons also show setting-specific failures and collapse.
Scope and limitations
The uncertainty estimator and its calibration are part of the assumptions. Avoiding a noisy-TV trap does not establish general causal discovery or reliable extrapolation.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.