Many modern combinatorial solvers have a variety of parameters through which a user can customise their behaviour. Algorithm configuration is the process of selecting good values for these parameters in order to improve performance. Time and again algorithm configuration has been shown to significa ntly improve the performance of many algorithms for solving challenging computational problems. Automated systems for tuning parameters regularly out-perform human experts, sometimes but orders of magnitude. Online algorithm configurators, such as ReACTR, are able to tune a solver online without incurring costly offline training. As such ReACTR’s main focus is on runtime minimisation while solving combinatorial problems. To do this ReACTR adopts a one-pass methodology where each instance in a stream of instances to be solved is considered only as it arrives. As such ReACTR’s performance is sensitive to the order in which instances arrive. It is still not understood which instance orderings positively or negatively effect the performance of ReACTR. This paper investigates the effect of instance ordering and grouping by empirically evaluating different instance orderings based on difficulty and feature values. Though the end use is generally unable to control the order in which instances arrive it is important to understand which orderings impact Re- ACTR’s performance and to what extent. This study also has practical benefit as such orderings can occur organically. For example as business grows the problems it may encounter, such as routing or scheduling, often grow in size and difficulty. ReACTR’s performance also depends strongly configuration selection procedure used. This component controls which configurations are selected to run in parallel from the internal configuration pool. This paper evaluates various ranking mechanisms and different ways of combining them to better understand how the candidate selection procedure affects realtime algorithm configuration. We show that certain selection procedures are superior to others and that the order which instances arrive in determines which selection procedure performs best. We find that both instance order and grouping can significantly affect the overall solving time of the online automatic algorithm configurator ReACTR. One of the more surprising discoveries is that having groupings of similar instances can actually negatively impact on the overall performance of the configurator. In particular we show that orderings based on nearly any instance feature values can lead to significant reductions in total runtime over random instance orderings. In addition, certain candidate selection procedures are more suited to certain orderings than others and selecting the correct one can show a marked improvement in solving times.

This paper outlines an online approach for algorithm configuration which uses the power of modern multicore system to evaluate multiple parameters configurations in parallel.

The success or failure of a solver is oftentimes closely tied to the proper configuration of the solver's parameters. However, tuning such parameters by hand requires expert knowledge, is time consuming, and is error-prone. In recent years, automatic algorithm configuration tools have made significant advances and can nearly always find better parameters than those found through hand tuning. However, current approaches require significant offline computational resources, and follow a train-once methodology that is unable to later adapt to changes in the type of problem solved. To this end, this paper presents Real-time Algorithm Configuration through Tournaments (ReACT), a method that does not require any offline training to perform algorithm configuration. ReACT exploits the multi-core infrastructure available on most modern machines to create a system that continuously searches for improving parameterizations, while guaranteeing a particular level of performance. The experimental results show that, despite the simplicity of the approach, ReACT quickly finds a set of parameters that is better than the default parameters and is competitive with state-of-the-art algorithm configurators.