Service interruption on Monday 11 July from 12:30 to 13:00: all the sites of the CCSD (HAL, EpiSciences, SciencesConf, AureHAL) will be inaccessible (network hardware connection).
Skip to Main content Skip to Navigation

Contributions to non-convex stochastic optimization and reinforcement learning

Abstract : This thesis is focused on the convergence analysis of some popular stochastic approximation methods in use in the machine learning community with applications to optimization and reinforcement learning.The first part of the thesis is devoted to a popular algorithm in deep learning called ADAM used for training neural networks. This variant of stochastic gradient descent is more generally useful for finding a local minimizer of a function. Assuming that the objective function is differentiable and non-convex, we establish the convergence of the iterates in the long run to the set of critical points under a stability condition in the constant stepsize regime. Then, we introduce a novel decreasing stepsize version of ADAM. Under mild assumptions, it is shown that the iterates are almost surely bounded and converge almost surely to critical points of the objective function. Finally, we analyze the fluctuations of the algorithm by means of a conditional central limit theorem.In the second part of the thesis, in the vanishing stepsizes regime, we generalize our convergence and fluctuations results to a stochastic optimization procedure unifying several variants of the stochastic gradient descent such as, among others, the stochastic heavy ball method, the Stochastic Nesterov Accelerated Gradient algorithm, and the widely used ADAM algorithm. We conclude this second part by an avoidance of traps result establishing the non-convergence of the general algorithm to undesired critical points, such as local maxima or saddle points. Here, the main ingredient is a new avoidance of traps result for non-autonomous settings, which is of independent interest.Finally, the last part of this thesis which is independent from the two previous parts, is concerned with the analysis of a stochastic approximation algorithm for reinforcement learning. In this last part, we propose an analysis of an online target-based actor-critic algorithm with linear function approximation in the discounted reward setting. Our algorithm uses three different timescales: one for the actor and two for the critic. Instead of using the standard single timescale temporal difference (TD) learning algorithm as a critic, we use a two timescales target-based version of TD learning closely inspired from practical actor-critic algorithms implementing target networks. First, we establish asymptotic convergence results for both the critic and the actor under Markovian sampling. Then, we provide a finite-time analysis showing the impact of incorporating a target network into actor-critic methods.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, December 17, 2021 - 10:10:01 AM
Last modification on : Saturday, December 18, 2021 - 3:17:43 AM
Long-term archiving on: : Friday, March 18, 2022 - 6:35:14 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03485159, version 1



Anas Barakat. Contributions to non-convex stochastic optimization and reinforcement learning. Machine Learning [stat.ML]. Institut Polytechnique de Paris, 2021. English. ⟨NNT : 2021IPPAT030⟩. ⟨tel-03485159⟩



Record views


Files downloads