Multiobjective Accelerated Gradient-like Flow with Asymptotic Vanishing Normalized Gradient

Abstract: In this paper, we extend the gradient system with unit-norm gradient term modification proposed by Wang et al.\cite{wang2021search} to multiobjective optimization, studying the following system: $$ \ddot x(t)+\frac{\alpha }{t}\dot x(t)+\frac{\alpha -\beta }{t^p}\frac{|\dot x(t)|}{|\proj_{C(x(t))}(0)|}\proj_{C(x(t))}(0)+\proj_{C(x(t))}(-\ddot x(t))=0 $$ where $C(x(t))=\textbf{conv}{\nabla f_i(x(t)):i=1,\cdots,m}$, $f_i(x(t)):\R^n\to \R$ are continuously differentiable convex functions, and $\alpha \ge \beta \ge 3$. Under certain assumptions, we establish the existence of trajectory solutions for this system. Using a merit function, we characterize the convergence of trajectory solutions: For $p>1$, $\alpha >\beta \ge 3$, we obtain a convergence rate of $O(1/t^2)$. When $\beta >3$, the trajectory solutions converge to a weak Pareto solution of the multiobjective optimization problem $\min _{x}(f_1(x),\cdots,f_m(x))^\top$. For $p=1$, $\alpha >\beta \ge 3$, we derive a convergence rate of $O(\ln^2 t/t^2)$. We further generalize Wang et al.'s FISC-nes algorithm to multiobjective optimization, achieving a convergence rate of $O(\ln^2k/k^2)$. The numerical experiments demonstrate that our system and algorithm exhibit competitive performance.