Global ETD Search
Search theses and dissertations gathered from participating repositories worldwide. Every result links back to the library that holds it. No account is needed.
Results
Showing 1 to 15 of 15 for “"Last-iterate"”.
-
Last iterate convergence in network zero-sum games
… in this generalization. We then generalize last iterate convergence results obtained previously in two-player zero-sum games. We analyze convergence rates when players update their strategies using Mirror Descent, and Optimistic Mirror Descent. We analyze these algorithms when they are used …
-
Singular Diffusion Limit of Tagged Particle Dynamics in a 1 Dimensional Zero Range Process With Sinai-Type Medium and the Convergence Rate for the Last Iterate of Stochastic Gradient Descent Schemes
… is to compute the convergence rate for the last iterate of stochastic gradient descent schemes. With respect to these problems, this thesis is similar in presentation to [HLS25] and [Hud25] respectively, but we augment with details, motivations, and ideas in important parts. At the time of …
-
Formal guarantees for heuristic optimization algorithms used in machine learning
… to noise under PL condition. Third, we study the last iterate of momentum methods. We prove the first lower bound in the general convex setting for the last iterate of SGD with constant momentum. Based on the fact that the lower bound is suboptimal, we investigate a class of (both adaptive and …
-
Convergence in min-max optimization
… by Karlin from 1959. We also show that linear last-iterate convergence rates are possible for the Hamiltonian Gradient Descent algorithm for the class of “sufficiently bilinear” min-max problems. Finally, we explore higher-order methods for min-max optimization and monotone variational …
-
Across the Reinforcement Learning and Large Language Model Spectrum
… online direct preference optimization (DPO) and last-iterate convergence guarantees for Nash learning under non-transitive preferences. The final part turns to RL for LLMs, presenting reflection-assisted online fine-tuning for agentic tasks, and a large-scale distributed RL post-training …
-
Provably effective algorithms for min-max optimization
… 2. Can projection-free algorithms achieve last-iterate convergence for constrained min-max optimization problems with the convex-concave landscape? 3. Can we show that stochastic gradient descent-ascent, a method commonly used in practice for GAN training, actually finds global optima and …
-
Learning and Optimization on Geodesic Metric Spaces
… to achieve sub-linear regret on manifolds. Lastly, we consider the last-iterate convergence of Riemannian extragradient-type methods, which can be provably employed to tackle Riemannian minimax problems. The result in Euclidean space, achieved by extragradient, is …
-
Enhancing optimization algorithms for non-convex settings : federated and decentralized learning
In the last decade, the rise and unprecedented success of neural networks have drawn significant attention from the machine learning community to non-convex optimization. This thesis focuses on federated and decentralized learning within the framework of non-convex optimization, aiming to deepen …
-
OFFLINE REINFORCEMENT LEARNING FOR MDPS AND CONFOUNDED POMDPS: POLICY GRADIENT ALGORITHMS AND PESSIMISTIC MODEL-BASED APPROACHES
… on the statistical error, and 3) a proof of last-iterate global convergence of the proposed algorithm. Lastly, we introduce a model-based offline RL method for confounded POMDPs under general function approximations, which is provably efficient under the assumption of partial coverage imposed …
-
Understanding the role of averaging in non-smooth stochastic gradient descent
… gradient descent (SGD), the error of the final iterate is O(logT⁄T) with high probability. We also construct a function from this class for which the error of the final iterate of deterministic gradient descent is Ω(logT⁄T). This shows that the upper bound is tight and that, in this setting, …
-
Nonparametric sparse learning of dynamical systems
… guarantees in the batch learning scenario and last-iterate convergence guarantees in the online case with streaming samples. To illustrate the efficacy of our sparse learning approach, we empirically examine the algorithm through deterministic and stochastic nonlinear system examples. We …
-
Provably efficient algorithms for decentralized optimization
… method with double averaging that guarantees the last iterate convergence, which is crucial to solving decentralized dual Lagrangian problems with convergence rate guarantee. Chapter 4 studies the composite cost-coupled decentralized optimization in stochastic networks, for which existing …
-
On Convergence Analysis of Stochastic and Distributed Gradient-Based Algorithms
… average of the squared gradient, as well as the last iterations of the algorithm converges to zero almost surely. We also provide a last-iterate almost sure convergence rate analysis for this method on general convex smooth functions. Another aspect of the research addresses the convergence rate …
-
Study on the effect of misinformation in multi-agent systems
… using best-response strategies, that guarantees last-iterate convergence for zero-sum games with a unique Nash equilibrium. Next, we consider the case of misinformation games where the misinformation is due to random noise that additively distorts the payoff matrices of the agents (e.g., due to …