Improved Parallel Algorithms for Sequential Minimal Optimization of Classification Problems
Wenjing Wei, LI Chang-rong, Jun Hai Guo · 2018
An approach presented in our previous work showed that sequential minimal optimization (SMO) algorithms could be executed in parallel. However, its convergence was not guaranteed in some cases. To solve the problem, we propose two novel algorithms in this paper. The first one is to add a condition to decide whether a single violating pair or multiple pairs should be chosen. Thus, the objective function decreases strictly in each iteration, and the algorithm is convergent. The second one aims at choosing better working set in each iteration to reduce the number of entire iterations. Due to the violating pairs chosen in each parallel way are violating at the updating moment, the convergence can be guaranteed. The results of our experiments show the two proposed algorithms can be executed successfully and their convergences are guaranteed totally.