Improving the Q rule-base adaptation in the HFRIQ-learning

Abstract The heuristically accelerated fuzzy rule interpolation-based Q-learning enables the integration and fine-tuning of external expert (a priori) knowledge during the learning. The system represents its knowledge base as a sparse fuzzy rule-base defined by a fuzzy rule interpolation Q-function. Expert knowledge is optimized through a gradient-based approach that requires computing the partial derivatives of the Q-function. The aim of this paper is to present the determination of these partial derivatives and describe analytical and numerical methods for their calculation. The results provide a foundation for gradient-based optimization in heuristically accelerated fuzzy rule interpolation-based Q-learning, enabling more precise and reliable refinement of expert knowledge.