VPFlow advances offline reinforcement learning with value-predictive planning

Researchers have developed VPFlow, a novel approach to offline goal-conditioned reinforcement learning that prioritizes value prediction over state prediction for planning. This method leverages a closed-form Monte Carlo return for value regression, training latent world models to predict goal-reaching value rather than future states. VPFlow utilizes these planning-oriented representations with a one-step flow policy and evolutionary search, demonstrating significant improvements in success rates on OGBench datasets compared to existing model-free and latent planners. AI

IMPACT This research advances offline reinforcement learning by demonstrating the efficacy of value prediction over state prediction for planning, potentially improving agent adaptability in complex tasks.

RANK_REASON The item describes a new research paper detailing a novel method for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• Flow policy
• Latent World Models
• LeWorldModel
• Monte Carlo return
• Offline Goal Reaching
• OGBench
• reinforcement learning
• TD-MPC2
• Value-Predictive Planning
• Value Regression
• VPFlow

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →