Learning a Go Heuristic with TILDE
CG '00 Revised Papers from the Second International Conference on Computers and Games
| Hi-index | 0.00 |
It has been said to be very useful for Go playing systems to have knowledge. We focus on pattern level knowledge and propose a new model of pattern acquisition based on our cognitive experiments. The model consists of two steps: pattern acquisition step, using only positive examples, and pattern refinement step, using both positive and negative examples. The latter step acquires precise conditions to apply and/or the way of conflict resolution. This model has advantages in computational time and precise control for conflict resolution. One algorithm is given for each step, and each algorithm can change independently, it is possible to compare algorithms with this model. Three algorithms are introduced for the first step and two for the second step. Patterns acquired by this model are applied to Tsume-Go problems (life and death problems) and the performance between six conditions are compared. In the best condition, the percentage of correct answers is about 31%. This result equals the achievement of one dan human players. It is also shown that the patterns enhance search techniques when the search space is very large.