Publications by authors named "Sokol Koco"
Greedy Methods, Randomization Approaches, and Multiarm Bandit Algorithms for Efficient Sparsity-Constrained Optimization.
IEEE Trans Neural Netw Learn Syst
November 2017
Article Synopsis
- The text discusses the limitations of traditional sparsity-constrained algorithms like OMP and FW, which can be computationally heavy due to the need to calculate gradient components for selecting non-zero variables.
- The authors propose innovations that focus on estimating just the top entry of the gradient using greedy methods and randomization techniques, thereby reducing computational costs.
- They introduce a bandit-based algorithm for efficiently identifying the best gradient entry and demonstrate through experiments that their methods can significantly speed up existing algorithms while maintaining similar performance to exact computations.