The　dynamic　modeling　for　magnetorheological　(MR)　dampers　to　describe　their　highly　nonlinear　dynamic　characteristics　is　essential　for　the　design　and　implementation　of　a　smart　MR　control　system.　One　critical　concern　in　constructing　a　nonparametric　MR　damper　model　by　employing　the　artificial　neural　network　technique　is　its　generalization　capability,　which　is　also　significant　to　guarantee　the　stability　and　reliability　of　the　MR　control　system.　The　paper　presents　the　modeling　of　MR　dampers　with　the　employment　of　the　NARX　(nonlinear　autoregressive　with　exogenous　inputs)　network　technique　within　a　Bayesian　inference　framework,　and　addresses　the　enhancement　of　model　prediction　accuracy　and　generalization　capability　in　terms　of　the　network　architecture　optimization　and　regularized　network　learning　algorithm.　The　Bayesian　regularized　NARX　network　model　for　the　MR　damper　is　demonstrated　to　outperform　the　non-regularized　network　model　with　the　superior　prediction　and　generalization　performance　in　the　scenarios　of　harmonic　and　random　excitations.　Therefore,　the　proposed　model　with　enhanced　generalization　is　beneficial　to　realize　the　real-time　and　robust　smart　control　of　MR　systems.　©　2017,　Editorial　Office　of　Journal　of　Vibration　and　Shock.　All　right　reserved.