1l.b introduce inrhis paper a new j-actahough set modeling nppronch lo the domains of atlraction of nonlinear oswteiiis ohtnined by cell mapping. The state space is pnrtitionect into cells and the stabilih regions found itsing cell io cell mapping. Our new approach gives a ~frnctnl.roiigli #set identi@ to the domains of attrnclion where cell.\. ore identified according to their ji-actnl dinrension ns ,fiilLv stable, possibly stnble and unstable. TheKe the srnbility domain is a rough set where fu& .rtable cells tlereriirine the lower approximation of the doninin, and posssihlv stable cells its rough hounday Consequentl~J. the totolip of these cells forms an upper approximation to tlie rough stability domain. The bountlarv of this doinnin which is a rough set of cells having a j-actnl (Iiiiierision as an attribute o f roughness is smoothed% minimising the inherent stability uncertain@ of the region, itsiiig n reinforcement learning technique whicli takes into nccount the stability historv of each fractnl:roiicgh cell. This new approach intended to reinfirce the perjbrrrrnnce of a controller under stability uncertnintl.! i s npplietl ,fiw illustrative purposes to a twoaxis robot nrm untler civiiamic load.