A learning algorithm that always learns best alternatives

A learning algorithm is described, and a probabilistic proof that it always learns to overwhelmingly prefer best alternatives, even in the presence of noise, is given. Learning is modeled as retention of memory: the learner learns to prefer the choice of alternative A in setting S because the learner remembers that the alternative often led to success in the past. The learning algorithm is illustrated in the context of a simple game (NIM) for which the rate and accuracy of learning can easily be measured. The algorithm can be seen as learning from examples, or even as learning by discovery through directed search. The searching is directed, since it increasingly reexamines and increasingly favors areas of past success. The algorithm solves the problem of orphan alternatives. It is, however, a weak method, and for it to be practical it must be restricted to domains where the number of alternatives being weighed is small. Experimental results are reported