This　paper　presents　an　online　architecture　design　algorithm　for　radical　basis　function　(RBF)　neural　network　based　on　online　subtractive　clustering　algorithm　aiming　at　designing　the　minimal　RBF　neural　network　structure.　The　algorithm　combines　the　characteristics　that　the　online　substractive　clustering　can　track　the　real-time　condition　with　the　parameters　learning　process　of　the　RBF　neural　network,　which　makes　the　RBF　neural　network　adapt　to　the　change　of　real-time　condition　dynamics　while　maintaining　a　compact　network　architecture.　Therefore,　this　method　can　effectively　solve　the　problem　of　self-organizing　structure　design　of　the　RBF　neural　network.　Only　the　kernel　function　whose　Euclidean　distance　is　nearest　to　the　real-time　conditions　is　adjusted,　which　greatly　improves　the　learning　speed　of　the　network.　The　results　of　experiments　on　the　typical　function　approximation　and　the　chaotic　time　series　prediction　show　that　the　algorithm　owns　favorable　dynamic　character　response　and　approximating　ability.