Tuning Strategies for Heterogeneous Instance Sets Using irace
TL;DR
An extension of the irace framework, h-irace, is proposed that incorporates concurrent racing processes over predefined instance subsets that supports configuration specialization across heterogeneous instance subsets in both continuous and combinatorial optimization problems.
Abstract
Automatic algorithm configuration is a well-established problem in optimization that becomes particularly challenging when dealing with heterogeneous sets of problem instances. In such scenarios, instance characteristics such as difficulty, structure, or dimensionality can induce substantial performance variability, leading to a trade-off between exploring diverse configurations for generalization and intensifying the search to improve performance. While specialized configurations for each instance type would be ideal, this is often unfeasible due to limited tuning budgets and the lack of reliable knowledge about instance similarity. To explore this issue, we propose an extension of the irace framework, h-irace, that incorporates concurrent racing processes over predefined instance subsets. The proposed approach is evaluated on two heterogeneous tuning scenarios using algorithms from the Ant Colony Optimization (ACO) and PSO-X families. The results show that the proposed method supports configuration specialization across heterogeneous instance subsets in both continuous and combinatorial optimization problems.