The authors study how temporal horizon affects feature learning in deep reinforcement learning.
Detached shorter-horizon features boost long-horizon reinforcement learning
Che, Ivanovic, and Pavone train separate smaller-discount value networks and feed their features through stop-gradients into a behavior learner whose discount and objective remain unchanged.
Big Tech
Tong Che · Boris Ivanovic · Marco Pavone
NVIDIA Research
Research Digest··2 min read
Che, Ivanovic, and Pavone propose Short-to-Long (S2L), where smaller-discount source value networks supply features to a behavior learner whose discount is unchanged.
§