State transition difference prediction for deep reinforcement learning

Haotian Chi

doi:10.1016/j.patcog.2025.112824

State transition difference prediction for deep reinforcement learning

2025-12-04

0

PRE

AI

DOI：10.1016/j.patcog.2025.112824

Original

Original request for help

Save

Abstract

En 中文

• We propose STDP, a novel RL method that learns latent dynamics by predicting transition differences between states. • STDP-E and STDP-I model explicit and implicit differences in state transitions, enabling complementary learning strategies. • STDP achieves superior sample efficiency over baselines on MuJoCo and D4RL in both online and offline RL tasks. • In-depth ablation and hyper-parameter analysis verify the effectiveness and robustness of STDP-E and STDP-I.

State transition difference prediction for deep reinforcement learning

Abstract

Journal

Organization