next up previous contents
Next: Koller and Pfeffer Up: Previous Computer Science Studies Previous: Findler

Machine learning

Waterman Waterman70 and Smith Smith83 used poker as a testbed for automatic learning methods. Both worked on the problem of acquiring problem-solving heuristics through experience. Waterman worked on two things, the representation of heuristics as rules to allow their dynamic manipulation, and the automatic modification and creation of these heuristics by a learning program based on information obtained during training. Waterman created a number of computer players differing in the heuristics initially provided to them during his research. The best program was evaluated to be the same degree of skill as a ``nonprofessional but experienced human player''. His programs played a two-player standard version of five-card draw poker and believed that by choosing poker, the representation and generalisation techniques he developed were shown to be an effective approach to implementing decision making under conditions of uncertainty and risk.

Smith proposed an alternative method for dynamically learning heuristics by using genetic algorithms. Poker was used as a testbed for this technique to provide a means for comparing his work with that of Waterman. The research done by Waterman and Smith is promising but was also not geared directly towards the development of a world class poker playing program. The use of learning algorithms appears to be of some use for the creation of a computer player, and is ideal for reasoning and learning in an imperfect information environment. It is uncertain how successful a program based on genetic algorithms or dynamic heuristic rules can become, but appears that it holds some potential and may be able to form the basis for future research into this area.


next up previous contents
Next: Koller and Pfeffer Up: Previous Computer Science Studies Previous: Findler
Jason R Carlton
2000-11-13