A unified approach to adaptive control of average reward Markov decision processes

The classical procedure for the adaptive control of average reward Markov decision processes with an unknown parameter chooses at each stage a decison which is optimal for the average reward problem with the presently estimated parameter. But in many cases it is inefficient or impossible to compute each time the long run optimal policy. So successive approximation methods were proposed and investigated. The paper presents a unifying and generalizing approach including both types of methods mentioned above and generating a lot of new procedures, too.