New SADQ method enhances Q-learning stability in Deep Q-Networks

Researchers have introduced the Successor Rollout Aggregation Deep Q-Network (SADQ), a novel modification to Q-learning designed to improve training stability in Deep Q-Networks (DQNs). SADQ addresses the issue of DQNs being overly sensitive to estimation noise by using one-step rollout predictions from a learned dynamics model. This approach guides the aggregation of future returns, reducing the amplification of errors caused by the maximization operator favoring unreliable estimates. Theoretical analysis and empirical results on various control tasks and Atari benchmarks demonstrate SADQ's effectiveness in mitigating overestimation and enhancing training stability compared to existing DQN variants. AI

IMPACT This research could lead to more stable and reliable training for reinforcement learning agents in complex environments.

RANK_REASON The cluster contains a research paper detailing a new method for Q-learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Atari
• DQNs
• Q-learning
• Successor Rollout Aggregation Deep Q-Network

• paper
• model release

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