Recently, Renault (2016) studied the dual bin packing problem in the per-request advice model of online algorithms. He showed that given $O(1/ε)$ advice bits for each input item allows approximating the dual bin packing problem online to within a factor of $1+ε$. Renault asked about the advice complexity of dual bin packing in the tape-advice model of online algorithms. We make progress on this question. Let $s$ be the maximum bit size of an input item weight. We present a conceptually simple online algorithm that with total advice $O\left(\frac{s + \log n}{ε^2}\right)$ approximates the dual bin packing to within a $1+ε$ factor. To this end, we describe and analyze a simple offline PTAS for the dual bin packing problem. Although a PTAS for a more general problem was known prior to our work (Kellerer 1999, Chekuri and Khanna 2006), our PTAS is arguably simpler to state and analyze. As a result, we could easily adapt our PTAS to obtain the advice-complexity result.
We also consider whether the dependence on $s$ is necessary in our algorithm. We show that if $s$ is unrestricted then for small enough $ε> 0$ obtaining a $1+ε$ approximation to the dual bin packing requires $Ω_ε(n)$ bits of advice. To establish this lower bound we analyze an online reduction that preserves the advice complexity and approximation ratio from the binary separation problem due to Boyar et al. (2016). We define two natural advice complexity classes that capture the distinction similar to the Turing machine world distinction between pseudo polynomial time algorithms and polynomial time algorithms. Our results on the dual bin packing problem imply the separation of the two classes in the advice complexity world.
Rényi-Ulam Games and Online Computation with Imperfect Advice
我们研究的新生设置的在线计算与不完美的建议,其中的在线算法是增强了一些预测编码的形式,可能是错误的二进制字符串。该算法不注意建议错误,但定义了期望的容限,即它可以容忍的错误建议位的数量的上限。这是一个推广不可信建议模型的模型[Angelopoulos et al. ITCS 2020],其中算法的性能仅在错误的极端值下进行评估(即,如果建议没有错误,或者如果它是逆向生成的)。在这项工作中,我们建立了一个说谎的回应者的游戏,也被称为Rényi-Ulam游戏之间的联系,并与不完美的建议在线算法的设计和分析。具体来说,我们演示了如何获得上界和下界的竞争力的比例,如时间序列搜索,在线招标,分数背包等研究在线问题。我们的技术提供了第一个下界在这个模型中的在线问题。我们还强调和利用竞争分析与不完美的建议和容错多处理器系统之间的连接。最后,我们展示了如何放弃对公差参数的依赖,通过资源增加和鲁棒化。
Stochastic Combinatorial Optimization via Poisson Approximation
研究了几种随机组合问题,包括期望效用最大化问题、随机背包问题和随机装箱问题。在这些问题中,一个常见的技术挑战是优化一组随机变量之和的某个函数。困难主要是由于和的概率分布是一组分布的卷积,这不是一个容易处理的目标函数。为了解决这个困难,我们引入泊松近似技术。该技术是基于泊松近似定理发现的Le Cam,它使我们能够近似分布的总和的一组随机变量使用复合泊松分布。我们首先研究最近引入的期望效用最大化问题[Li and Despande,FOCS11]。对于单调和Lipschitz效用函数,我们得到一个加性PTAS,如果有一个多维PTAS的多目标版本的问题,严格推广了以前的结果。对于随机装箱问题(在[Kleinberg、Rabani和Tardos,STOC 97]中介绍),我们证明了存在一个多项式时间算法,如果我们放宽每个箱的大小和溢出概率,则该算法最多使用最佳箱数eps。对于随机背包,我们展示了一个1+ eps近似使用eps额外容量,即使每个项目的大小和奖励可能是相关的,并允许取消项目。这概括了以前的工作[Balghat,Goel和Khanna,SODA 11]的情况下没有相关性和取消。我们的算法也更简单。我们还提出了一个因子2+eps近似算法的随机背包取消。目前已知的近似因子为8 [Gupta,Krishnaswamy,Molinaro and Ravi,FOCS11]。