On-Line SystemIdentification UsingContext Discernment
R.A. Santiago · 2005
Mathematical models areoften usedinsystem identification applications. Thedynamics ofmostsys- tems,however, change overtimeandthesources ofthese changes cannotalways bedirectly determined ormeas- ured. Tomaintain modelaccuracy, itisdesirable tode- sign systemidentifiers thatcanadapt tothese dynamical shifts. Weusereinforcement learning totrain anagentto recognize dynamical changes inamodeled system andto estimate newparameter values forthemodel. Thesubse- quentactions ofthis agentarecharacterized asmoving theparameterized modelon an optimal trajectory in modelparameter space. Itisfoundthatthis methodis capable ofquickly andaccurately discerning thecorrect parameter values.