Reducing the Seesaw Effect with Deep Proof-Number Search
| Autor: | Taichi Ishitobi, H. Jaap van den Herik, Hiroyuki Iida, Aske Plaat |
|---|---|
| Rok vydání: | 2015 |
| Předmět: | |
| Zdroj: | Lecture Notes in Computer Science ISBN: 9783319279916 ACG |
| DOI: | 10.1007/978-3-319-27992-3_17 |
| Popis: | In this paper, DeepPN is introduced. It is a modified version of PN-search. It introduces a procedure to solve the seesaw effect. DeepPN employs two important values associated with each node, viz. the usual proof number and a deep value. The deep value of a node is defined as the depth to which each child node has been searched. So, the deep value of a node shows the progress of the search in the depth direction. By mixing the proof numbers and the deep value, DeepPN works with two characteristics, viz., the best-first manner of search (equal to the original proof-number search) and the depth-first manner. By adjusting a parameter (called R in this paper) we can choose between best-first or depth-first behavior. In our experiments, we tried to find a balance between both manners of searching. As it turned out, best results were obtained at an R value in between the two extremes of best-first search (original proof number search) and depth-first search. Our experiments showed better results for DeepPN compared to the original PN-search: a point in between best-first and depth-first performed best. For random Othello and Hex positions, DeepPN works almost twice as good as PN-search. From the results, we may conclude that Deep Proof-Number Search outperforms PN-search considerably in Othello and Hex. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|