6533b873fe1ef96bd12d4aef

RESEARCH PRODUCT

An adaption mechanism for the error threshold of XCSF

subject

description

Learning Classifier System (LCS) is a class of rule-based learning algorithms, which combine reinforcement learning (RL) and genetic algorithm (GA) techniques to evolve a population of classifiers. The most prominent example is XCS, for which many variants have been proposed in the past, including XCSF for function approximation. Although XCSF is a promising candidate for supporting autonomy in computing systems, it still must undergo parameter optimization prior to deployment. However, in case the later deployment environment is unknown, a-priori parameter optimization is not possible, raising the need for XCSF to automatically determine suitable parameter values at run-time. One of the most important parameters is the error threshold, which can be interpreted as a target bound on the approximation error and has to be set according to the approximated function. To enable XCSF to reliably approximate functions unknown at design-time, we propose the use of an error threshold, which is adapted at run-time based on the currently achieved approximation error. Our experimental evaluation shows that the adaption mechanism automatically adjusts the error threshold to a suitable value. On six different target functions, XCSF with an adaptive error threshold leads to a worst-case approximation error that is 11.5% higher than the error achieved with the best-suited static threshold.