Many learning systems search through a space of possible performance elements,<br />seeking an element whose expected utility, over the distribution of problems, is<br />high. As the task of finding the globally optimal element is often intractable, many<br />practical learning systems instead hill-climb to a local optimum. Unfortunately,<br />even this is problematic as the learner typically does not know the underlying<br />distribution of problems, which it needs to determine an element's expected utility.<br />This paper addresses the task of approximating this hill-climbing search when the<br />utility function can only be estimated by sampling. We present a general algorithm,<br />palo, that returns an element that is, with provably high probability, essentially a<br />local optimum. We then demonstrate the generality of this algorithm by presenting<br />three distinct applications that respectively find an element whose efficiency,<br />accuracy or completeness is nearly optimal. These results suggest approaches to<br />solving the utility problem from explanation-based learning, the multiple extension<br />problem from nonmonotonic reasoning and the tractability/completeness tradeoff<br />problem from knowledge representation.
Greiner, R. (1996). PALO: a probabilistic hill-climbing algorithm. Artificial Intelligence, 84(1–2), 177–208. https://doi.org/10.1016/0004-3702(95)00040-2