Percentile optimization in uncertain Markov decision processes with application to efficient exploration