Research map / Learning theory and algorithms / Prediction and decision losses

A Surrogate Perspective on Convergence of Fixed-Target DQN

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Loss landscape analysis

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Analyzes fixed-target DQN convergence by treating the inner regression loop as minimizing a smooth surrogate loss whose progress is measured differently from the Bellman operator's gamma-contraction in l-infinity.

Problem

Fixed-target DQN relies on a Bellman optimality contraction in l-infinity norm, yet its inner regression minimizes smooth surrogate losses (MSE/Huber) measured in a different geometry, creating a mismatch in the convergence theory.

Approach

Views fixed-target DQN through a surrogate-loss lens, relating the inner regression progress to the Bellman optimality operator's contraction to explain convergence of the outer target-update loop.

Experimental setup

Builds on: DQN fixed-target value iteration

Closest papers by meaning

• On the Convergence and Sample Complexity Analysis of Deep Q-Networks with $\epsilon$-Greedy Exploration (NeurIPS 2023 · similarity 0.90)
• Universal Approximation Theorem of Deep Q-Networks (ICML 2025 · similarity 0.89)
• Convergent and Efficient Deep Q Learning Algorithm (ICLR 2022 · similarity 0.87)
• Towards Optimal Adversarial Robust Q-learning with Bellman Infinity-error (ICML 2024 · similarity 0.87)
• Solving hidden monotone variational inequalities with surrogate losses (ICLR 2025 · similarity 0.86)
• Layerwise LQR for Geometry-Aware Optimization of Deep Networks (NeurIPS 2026 · similarity 0.85)