This　paper　proposes　a　two-phase　identification　approach　to　Mamdani　fuzzy　neural　networks.　The　first　phase　is　the　system　identification　which　includes　a　novel　forward　recursive　input-output　clustering　method　for　the　structure　initialization　and　the　gradient　descent　algorithm　for　the　parameter　initialization.　The　main　advantage　of　the　proposed　method　is　that　it　fits　perfectly　the　special　clustering　requirement　for　system　identification:　coarser　clustering　in　the　regions　where　the　identified　system　is　smoother　and　finer　clustering　in　the　regions　where　the　system　is　more　variable　or　nonlinear.　The　second　phase　is　the　system　simplification　which　includes　the　accurate　similarity　analysis　and　merging　method　for　similar　fuzzy　rules　and　the　gradient　descent　algorithm　for　the　parameter　finalization.　The　accurate　similarity　analysis　developed　solves　the　long　standing　open　problem　how　to　compute　the　exact　(rather　than　approximate)　similarity　between　fuzzy　sets　and　rules　with　Gaussian　membership　functions.　Numerical　experiments　based　on　well-known　benchmark　data　sets　are　used　to　verify　the　effectiveness　and　accuracy　of　the　proposed　approach.　(C)　2016　Elsevier　B.V.　All　rights　reserved.