SMDP基于性能势的神经元动态规划

？Analpha-uniformizedMarkovchainisdefinedbytheconceptofequivalentinfinitesimalgeneratorforasemi-Markovdecisionprocess(SMDP)withbothaverage-anddiscounted-criteria.Accordingtotherelationsoftheirperformancemeasuresandperformancepotentials,theoptimizationofanSMDPcanberealizedbysimulatingthechain.Forthecriticmodelofneuro-dynamicprogramming(NDP),aneuro-policyiteration(NPI)algorithmispresented,andtheperformanceerrorboundisshownasthereareapproximateerrorandimprovementerrorineachiterationstep.TheobtainedresultsmaybeextendedtoMarkovsystems,andhavemuchapplicability.Finally,anumericalexampleisprovided.