This　paper　presents　the　development　and　application　of　a　soft-computing　technique　in　identification　of　forward　and　inverse　dynamics　of　a　self-sensing　magnetorheological　(MR)　damper　based　on　experimental　measurements.　This　technique　is　developed　by　the　synthesis　of　an　NARX　(nonlinear　autoregressive　with　exogenous　inputs)　model　structure　and　neural　network　within　a　Bayesian　inference　framework.　The　Bayesian　inference　procedures　essentially　eschew　overfitting　that　could　occur　in　network　learning　and　improve　generalization　(prediction)　capability　by　regularizing　the　complexity　of　learning.　In　applying　the　developed　technique　to　the　self-sensing　MR　damper,　the　present　and　past　information　of　its　input　and　output　quantities,　which　contain　the　physical　knowledge　of　the　damper,　is　used　to　formulate　its　nonlinear　dynamics.　The　NARX　network　architecture　is　then　optimized　to　enhance　modeling　effectiveness,　efficiency,　and　robustness.　Experimental　data　of　the　damper　subjected　to　both　harmonic　and　random　excitations　are　used　for　model　identification　and　assessment.　Assessment　results　show　that　the　formulated　Bayesian　NARX　network　accurately　emulates　the　nonlinear　forward　dynamics　of　the　self-sensing　MR　damper.　Improved　generalization　(prediction)　capability　of　the　NARX　network　model　by　the　Bayesian　regulation　is　observed　by　comparing　the　modeling　results　with　and　without　considering　regularization.　An　inverse　dynamic　model　for　the　self-sensing　MR　damper　is　further　formulated　by　the　developed　technique.　The　proposed　soft-computing　technique　is　viable　to　formulate　dynamic　models　of　the　self-sensing　MR　damper　for　structural　control　applications.　(C)　2015　American　Society　of　Civil　Engineers.