Back to feed
arXiv cs.LG
arXiv cs.LG
7/20/2026
Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control

Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control

Short summary

This paper improves the convergence rate of stochastic multi-gradient descent (MGDA) for multi-objective learning by analyzing the Hölder continuity of conflict-avoidant update directions. The authors show the CA direction is 1/2-Hölder continuous in the worst case but becomes Lipschitz under regularity conditions, enabling their proposed MoRe method to improve convergence from O(T^{-1/4}) to O(T^{-1/2}). The algorithm switches between CA direction updates and linear scalarization based on gradient conflict magnitude.

  • •CA direction is 1/2-Hölder continuous in worst case but Lipschitz under regularity conditions
  • •Proposed MoRe method improves stochastic MGDA convergence from O(T^{-1/4}) to O(T^{-1/2})
  • •Algorithm adapts between conflict-avoidant updates and linear scalarization based on gradient conflict

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv cs.LG

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Deep Reinforcement Learning for Reliability Based Bi-Objective Portfolio Optimization
arXiv cs.LG
Reinforcement Learning-Guided NSGA-II Enhanced with Gray Relational Coefficient for Multi-Objective Optimization: Application to NASDAQ Portfolio Optimization
arXiv cs.LG
Orthogonal Gradient Constraints Shape Noisy-Label Memorization Dynamics
arXiv cs.LG
Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization
arXiv cs.LG