Statistical comparative between selection rules for adaptive operator selection in vehicle routing and multi-knapsack problems

Abstract

Autonomous Search is an important field of artificial intelligence which focuses on the analysis and design of auto-adaptive systems capable to improving their own search performance on a given problem at runtime. An autonomous search system must can modify or select its internal components, improving its performance at execution time. A case of such algorithms is the Adaptive Operator Selection (AOS) method. This method uses a record of the latest iterations to propose which operator use in later iterations. AOS has two phases: Credit assignment and Selection rule. The first phase penalizes or rewards a specific operator based on their observed performance. The second phase makes the selection of the operator to use in subsequent iterations. This article shows the performance-based statistical comparison between two selection rules: Probability Matching and Adaptive Pursuit when using two different domains; namely, Vehicle Routing Problem and Multi-knap-Sack. The comparison is done with statistical rigor when integrating contrast algorithms such as no adaptation rule and random adaptation rule. This paper can be seen as an introduction on how to make proper statistical comparisons between two AOS rules.