In　this　paper,　a　novel　noise-constrained　least-squares　(NCLS)　method　for　online　autoregressive　(AR)　parameter　estimation　is　developed　under　blind　Gaussian　noise　environments,　and　a　discrete-time　learning　algorithm　with　a　fixed　step　length　is　proposed.　It　is　shown　that　the　proposed　learning　algorithm　converges　globally　to　an　AR　optimal　estimate.　Compared　with　conventional　second-order　and　high-order　statistical　algorithms,　the　proposed　learning　algorithm　can　obtain　a　robust　estimate　which　has　a　smaller　mean-square　error　than　the　conventional　least-squares　estimate.　Compared　with　the　learning　algorithm　based　on　the　generalized　least　absolute　deviation　method,　instead　of　minimizing　a　non-smooth　linear　L1　function,　the　proposed　learning　algorithm　minimizes　a　quadratic　convex　function　and　thus　is　suitable　for　online　parameter　estimation.　Simulation　results　confirm　that　the　proposed　learning　algorithm　can　obtain　more　accurate　estimates　with　a　fast　convergence　speed.　©　2009　Elsevier　Ltd.　All　rights　reserved.