Package	com.lookbackon.AI.searching
Class	public class AlphaBeta
Inheritance	AlphaBeta SearchingBase AbstractProcess Object

Search enhancements All techniques above aimed at reducing the number of nodes to search by better move ordering. There is another class of enhancements with the same goal, but with different means. These enhancements try to exploit the nature of the AlphaBeta algorithm, which has to search fewer nodes when the alpha-beta window is smaller. Note how AlphaBeta receives the parameters alpha and beta which tell it what range the value of the current position should lie. Once a move has returned with a higher value than alpha, this best value is saved in the variable localalpha and used for the next recursive call of AlphaBeta. If the best value is larger than beta, the search terminates immediately - we have found a move which refutes the notion that this position has a value in the range from alpha to beta, and do not need to look for another one. Note how my AlphaBeta function is returning the highest value it found, this can be higher than beta. Some people prefer to return beta instead of the best value on a fail high, that formulation of AlphaBeta is known as fail-hard. My formulation above is called fail-soft. The names come from the fact that in fail hard, the bounds alpha and beta are "hard", the return value cannot be outside the alpha-beta window. It would seem that fail-soft is much more sensible, as it might lead to more cutoffs: If you can return a higher value than beta (or a lower value than alpha), then perhaps you might get a cutoff in a previous instance of AlphBeta at a lower level in the search tree that you wouldn't get otherwise. However, the fail-hard camp says they get less search instabilities when using advanced techniques such as pruning.

See also

---