Action selection for stochastic, delayed reward

The paper gives a novel account of quick decision making for maximising delayed reward in a stochastic world. The approach rests on observable operator models of stochastic systems, which generalize hidden Markov models. A particular kind of decision situations is outlined, and an algorithm is presented which allows to estimate the probability of future reward with a computational cost of only O(im), where i is the number of action alternatives and m is the model dimension.