Generic approaches for parallel rule matching in learning classifier systems

The XCS classifier system (XCS) constitutes the most deeply investigated evolutionary rule-based machine learning algorithm. Due to its online learning nature, it is not as computationally intense as deep learning approaches. However, in order to increase XCS's realtime capabilities, it is worth to improve its runtime for the inference step. A iteration through XCS's main loop involves various steps. Among them, matching usually constitutes the computationally most expensive part. Various ways for improvements have been proposed, e.g. using specialized hardware or vector commands. However, existing approaches require either a certain hardware coupled with a specific instruction set, or a dedicated multiprocessing programming language. In this context, parallel programs are often evaluated in terms of their speed up, but it turns out that qualitative criteria such as flexibility and independence of a specific hardware are also relevant. We therefore present a generic parallel approach for matching. We only rely on the minimum assumption of a multicore CPU with standard synchronization mechanisms, e.g., locks. Those assumptions are satisfied in almost all modern computing systems and high-level programming languages today. Thus we show that our approaches cannot only decrease the runtime, but are also reusable for most learning classifier systems and computing systems.