An Adaptive Granular-Ball Generation Method Based on Reinforcement Learning

The granular-ball computing model uses an adaptive multi-granularity data representation for clustering, classification, and graph computing. We propose an adaptive granular-ball generation model based on reinforcement learning. Compact clustering initialization is followed by policy-guided splitting and centroid movement, with a reward based on purity, density, overlap, and ball count. Experiments on five image benchmarks characterize the balance between representation size and purity. Additional comparisons show reduced generation time relative to ORIGIN and a compact nearest-centroid representation; the CIFAR-10 case combines higher reported classification accuracy with fewer stored references than both baselines.