We consider online submodular minimisation under bandit feedback, where at each round a decision maker plays a subset and observes only its cost, while the cost functions, and hence their minimiser, change over time. Existing bandit methods only guarantee static regret, which compares against a single fixed subset and is a weak benchmark when the minimiser moves. This paper provides, to the best of our knowledge, the first method with guarantees for tracking a moving minimiser in this setting. We propose a zeroth-order method built on a two-point oracle for the Lovász extension whose two evaluations are both subsets that can be played. The perturbed and base points are rounded by a common random threshold, and drawing the smoothing direction from the unit sphere makes the base evaluation a zero-mean control variate, so the oracle is an unbiased estimator of the smoothed gradient. We establish dynamic regret bounds of order $O\big((1+\sqrt{P^\ast_N})N^{3/4}\big)$ in general and $O\big((1+\sqrt{P^\ast_N})N^{2/3}\big)$ under a sufficiently small drift, where $P^\ast_N$ is the path length of the minimiser sequence, together with the corresponding $O(N^{3/4})$ and $O(N^{2/3})$ static regret bounds. Numerically, the method attains lower regret than the algorithm of (Hazan and Kale (2012)), with the advantage growing with the path length.
Amir Ali Farzin, Yuen-Man Pun, Philipp Braun, Iman Shames
原文摘要
We establish a connection between the stability of discrete-time algorithms (DTAs) and continuous-time dynamical systems, in particular their $O(s^r)$-resolution ordinary differential equations (ODEs). We show that exponential stability of a common equilibrium transfers to the DTA whenever the one-step error is small relative to the distance to the equilibrium and the one-step Lyapunov decay. This condition is sharp in order, and one-step consistency alone is insufficient. The result is local and allows step-size-dependent ODEs and Lyapunov functions, covering $O(s)$-resolution ODEs whose decay rate vanishes as the step size tends to zero. Asymptotic stability is not inherited in general, but a comparison criterion and a checkable compatibility condition ensure it carries over for equilibria and compact sets. For resolution ODEs, joint regularity of the update map guarantees the required error bound. We apply the framework to two-timescale gradient descent ascent (TT-GDA), generalised extragradient, two-timescale proximal point, and (regularised) damped Newton, showing that for small step sizes and suitable parameters, every saddle point with invertible Hessian is locally exponentially stable for all but TT-GDA. For degenerate saddle points, the framework yields asymptotic stability and convergence rates.