With deep regret I found it impossible to avoid the use of game progress to add weights to evaluation features that are more important in the endgame than in the opening or the middle game. Penalty for blocked pawns, knight and silvers.Balance of attack cover and defense cover on the 25 squares around the king.Assault value of pieces (for example, how to play a climbing silver attack).Castle value of pieces (for example, where to have the golds in a yagura castle).Defense distance of single pieces to the king.Attack distance of single pieces to the king.Therefore, the evaluation function of my program only evaluated the following things: As a result, the evaluation function had to be simple enough to tune by hand or it would be impossible to make a stable program. Using automatic learning is in vogue since Hoki published the machine learning technique he used in the 2006 World Champion program Bonanza, but rather than copying this method I want to use my program for other ways of using knowledge and I decided not to implement Bonanza's learning at this point. I also made a conscious effort to keep the evaluation function as simple as possible. Not close to the top programs that do about 10,000,000 positions on 8 core machines, but still quite good. This gave the program another important performance boost, and running on a 4 core machine it searched about 2,000,000 positions per second on average. Because I was already using an implementation close to Crafty, it was not difficult to add this and I had parallelism working in about two weeks, although I still found some minor bugs in the weeks after that. The main addition to the 2008 Spear was the implementation of running the search in parallel. In the 2007 tournament I had entered with a program based on bitboards, which seems to be a good starting point, even in games where the number of squares is not a power of two. In the past I had entered the tournament with unstable programs that were hanging together with ad-hoc solutions, but it turned out that it was almost impossible to improve these programs and sooner or later they had to be discarded. When I rebuilt my program Spear from scratch in the summer of 2006, I was aiming for a program with solid foundations, which would be easier to improve. Reijer is now directlyĬontributing to our latest Shogi program, due for market later this year. Programs have been developed in parallel. With the development of AI Factory's Shogi engine Shotest, as these two Reijer is a lecturer at the Department of Informatics at Yamagata UniversityĪnd the author of the shogi program SPEAR, developed over the last 10 yearsĪnd currently in its third re-write.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. Archives
December 2022
Categories |