Network slicing offers the potential to enhance service satisfaction and optimize resource utilization, particularly in scenarios where radio resources are limited. Subslicing has been shown to improve the slice performance. In this paper, the monitor-analyze-plan-execute-knowledge (MAPE-K)-type management closed control loop (MCCL) is implemented for slice performance improvement by subslicing. The subslicing can improve slice performance if the slice performance depends on the size of slice bandwidth part (BWP). For Plan function, the classifier neural network was trained to decide whether the subslice should be split, merged or not changed by their performance. The training data contains slice performance data of all possible subslice sizes. For Execute function, the subslice splitting algorithm was proposed, which clusters UEs by their block error ratio (BLER) and allocates bandwidth proportionally to group requested sum rate and group BLER. A realistic 5G new radio (NR) band serving a set of user equipments (UE) of different values of their BLER and requested rates was a setup of radio access network (RAN) slice simulated using MATLAB R2021b. Subslicing has reduced bandwidth utilization, and slice BLER while increased slice goodput (application-level throughput). Proposed subslice splitting algorithm when UEs are clustered by their achieved BLER, then the slice BLER reduces additional 20% and slice goodput increases up to additional 9% compared to no subslicing when UEs are clustered by their requested rates. This effect was larger for the uplink. In runtime scenarios for poor-BLER UEs the smaller subslices improve slice utilization and BLER, while larger subslices improve goodput.