arxivcs.LG2026-07-04
Memory Merge DQN: Sensitivity Weighted Target Updates for Stable Value Learning
Adrian Ly, Richard Dazeley, Peter Vamplew, Sunil Aryal, Francisco Cruz
Deep Q-networks use target networks to stabilise bootstrapped value learning, but the standard hard copy update also introduces a tradeoff. Holding the target network fixed, improves short term stability, yet each hard update abruptly replaces the target parameters with the newes…