Distributed and Parallel ADMM for Structured Nonconvex Optimization Problem.

Area of Science:

• Computational Mathematics
• Optimization Theory
• Machine Learning

Background:

• Nonconvex optimization problems are increasingly important but lack efficient algorithms and robust theory.
• Structured, large-scale nonconvex problems pose significant challenges in real-world applications.
• Existing methods struggle with high dimensionality and distributed data.

Purpose of the Study:

• To propose an application-driven algorithmic framework for structured nonconvex optimization.
• To address challenges of high dimensionality and distributed training data simultaneously.
• To establish theoretical convergence guarantees for the proposed algorithm.

Main Methods:

• Development of a novel algorithmic framework utilizing distributed and parallel computing techniques.
• Jointly handling high dimensionality of model parameters and distributed training datasets.
• Theoretical analysis to establish convergence under moderate assumptions.

Main Results:

• The proposed algorithm demonstrates theoretical convergence under moderate assumptions.
• Successful application to popular multitask learning problems, including reinforcement learning.
• Promising performance indicating effectiveness and efficiency of the framework.

Conclusions:

• The developed framework offers an effective and efficient solution for structured nonconvex optimization.
• The approach is suitable for large-scale problems with distributed data.
• The method shows potential for various multitask applications.