Portal de Periódicos da CAPES

Machine learning for optimal blackjack counting strategies

1992; Elsevier BV; Volume: 33; Issue: 3 Linguagem: Inglês

10.1016/0378-3758(92)90001-9

1873-1171

S. Yakowitz, M. Kollier,

Machine Learning and Data Classification

Search for optimal counting strategies for blackjack, like many other tasks in artificial intelligence, can be transcribed into the task of finding a global minimum of a multi-modal discontinuous objective function on the basis of noisy measurements. Herein we relate an algorithm for such ‘stochastic minimization’ problems, and derive some general properties. The general framework is extended to be applicable to finding the ‘ten-count’ standing-number strategy for blackjack. Computational experiments explore this learning application. The feature here, as opposed to blackjack analyses by Thorp, Braum, and others, who rely on combinatorics in their constructions, is that our methodology requires minimal modelling or understanding of the problem structure. This is the first reasonably detailed account of a procedure for finding ten-count standing numbers. But it is general enough to be applicable to other blackjack problems and, in fact, to a wide variety of sequential decision tasks.