Slow features from goal-directed Markov chains flatten near high-occupancy states, hurting value-function approximation; rescaling features by the square root of the stationary distribution largely repairs this, while goal-averse behavior can even help.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Slow Feature Analysis on Markov Chains from Goal-Directed Behavior
Slow features from goal-directed Markov chains flatten near high-occupancy states, hurting value-function approximation; rescaling features by the square root of the stationary distribution largely repairs this, while goal-averse behavior can even help.