Multi-criteria Reinforcement Learning

Zoltán Gábor, Zsolt Kalmár, Csaba Szepesvári

Abstract: Ve cOllf:iider multi- criteria f:iequent,ial decision making problems where the vcctor-"valucd evaluations arc compared by a given, fixed total ordering. Condit.ions for the opt.irnality of statiOIl<-l,r}' p()lichs ;-weI the Bellman opti­ malit,y equatio n a re given for a. speci al, but. important class of problems ''v hell the eval­ uation of policies can be computed for the criteria, independently of each other. The anal)'sis requires special cafC as t.he t.opol­ ag,Y int.roduced by polnL\visc convergence a.ncl the or<1cr- topology introduced by the prefer­ ence order arc in general incompatible. Re­ inf orcement. learning algorithms are proposed and analY7,ed.