arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1609.08262math.OC
收藏

分布式正则化原-对偶方法:收敛性分析与折衷

Distributed Regularized Primal-Dual Method: Convergence Analysis and Trade-offs

Masoud Badiei Khuzani, Na Li

摘要:

研究了可分目标函数全局不等式约束分布式优化问题的确定性和随机原始-对偶次梯度算法。在这两种算法中,拉格朗日乘子的范数通过用二次正则化项增广相应的拉格朗日函数来控制。具体来说,我们表明,当每个算法的步长满足一定的限制,拉格朗日乘子的范数是上界的表达式,是成反比的正则化参数。我们使用这个结果来计算拉格朗日函数的次梯度的上界。对于确定性算法,我们证明了达到最优目标值的收敛速度。在随机最优化的情况下,我们同样证明了收敛速度的期望和高概率,使用有界鞅差的方法。对于这两种算法,我们证明了一个权衡之间的收敛速度和约束违反的衰减速率,在这个意义上说,提高收敛速度减缓约束违反的衰减速率,反之亦然。我们证明了我们提出的算法的收敛性数值分布回归铰链和逻辑损失函数在不同的图形结构。

英文摘要:

We study deterministic and stochastic primal-dual sub-gradient algorithms for distributed optimization of a separable objective function with global inequality constraints. In both algorithms, the norm of the Lagrangian multipliers are controlled by augmenting the corresponding Lagrangian function with a quadratic regularization term. Specifically, we show that when the stepsize of each algorithm satisfies a certain restriction, the norm of the Lagrangian multipliers is upper bounded by an expression that is inversely proportional to the parameter of the regularization. We use this result to compute upper bounds on the sub-gradients of the Lagrangian function. For the deterministic algorithm, we prove a convergence rate for attaining the optimal objective value. In the stochastic optimization case, we similarly prove convergence rates both in the expectation and with a high probability, using the method of bounded martingale difference. For both algorithms, we demonstrate a trade-off between the convergence rate and the decay rate of the constraint violation, in the sense that improving the convergence rate slows the decay rate of the constraint violation and vice versa. We demonstrate the convergence of our proposed algorithms numerically for distributed regression with the hinge and logistic loss functions over different graph structures.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑