Minimizing Dynamic Regret and Adaptive Regret Concurrently
Authors: Lijun Zhang, Shiyin Lu, Tianbao Yang
Abstract: Regret minimization is dealt with as a result of the golden rule inside the standard look at of on-line learning. However, regret minimization algorithms are inclined to converge to the static optimum, thus being suboptimal for altering environments. To cope with this limitation, new effectivity measures, along with dynamic regret and adaptive regret have been proposed to data the design of on-line algorithms. The earlier one objectives to cut back the worldwide regret with respect to a sequence of adjusting comparators, and the latter one makes an try to cut back every native regret with respect to a tough and quick comparator. Present algorithms for dynamic regret and adaptive regret are developed independently, and solely purpose one effectivity measure. On this paper, we bridge this gap by proposing novel on-line algorithms which may be able to lower the dynamic regret and adaptive regret concurrently. Truly, our theoretical guarantee is even stronger inside the sense that one algorithm is able to lower the dynamic regret over any interval
