With the goal of devising algorithms for decision support in operational tasks, we introduce a new methodology for the automated configuration of algorithms for combinatorial online optimization problems. The procedure draws upon available instance data and is capable of recognizing data patterns which prove beneficial to the overall outcome. Since online optimization requires repetitive decision making without complete future information, no online algorithm can be optimal for every instance and it is reasonable to restrict attention to rule-based algorithms. We consider such algorithms in the form of procedures which derive their decisions using a threshold value. Threshold values are computed by evaluating a mathematical term (threshold value expression) composed of the available instance data elements. The goal then consists of determining the structure of the threshold value expression leading to the best algorithm performance. To this end, we employ a simulated annealing scheme returning the most favorable term composition given the available instance data. The resulting methodology can be implemented as part of data-driven decision support systems in order to facilitate knowledge-based decision making. Decision rules are generated in an automated fashion once historical input data is provided. The methodology is successfully instantiated in a series of computational experiments for three classes of combinatorial online optimization problems (scheduling, packing, lot sizing). Results show that automatically configured online algorithms are even capable of substantially outperforming well-known online algorithms in respective problem settings. We attribute this effect to the methodology's capability of integrating instance data into the process of algorithm configuration.

为了设计用于操作任务中的决策支持的算法，我们引入了一种用于组合在线优化问题的算法自动配置的新方法。该过程利用可用的实例数据，并且能够识别证明对总体结果有益的数据模式。由于在线优化需要重复的决策而没有完整的未来信息，因此没有一种在线算法可以针对每个实例都是最优的，因此将注意力集中在基于规则的算法上是合理的。我们以过程的形式考虑这些算法，这些过程使用阈值得出其决策。通过评估由可用实例数据元素组成的数学项（阈值表达式）来计算阈值。然后，目标包括确定导致最佳算法性能的阈值表达式的结构。为此，我们使用模拟退火方案，在给出可用实例数据的情况下返回最有利的术语组合。可以将所得方法论作为数据驱动的决策支持系统的一部分实施，以促进基于知识的决策制定。提供历史输入数据后，决策规则将以自动方式生成。该方法已在针对三类组合在线优化问题（计划，包装，批量确定）的一系列计算实验中成功实例化。结果表明，在各个问题设置中，自动配置的在线算法甚至甚至能够大大胜过众所周知的在线算法。