SnailChess is a Windows program written in C and compiled with MS Visual C for the Pentium Compatible processor. It does not have its own graphical interface and requires to run with the chess playing interface Winboard by Tim Mann.

The basic of the program follows mainly what is learned from the codes of TSCP by Tom Kerrigan and that of Gerbil by Bruce Moreland. Both programs are very good sources for beginners.

The author started chess programming in 1997 and has released previous versions. This version 4.0 is a new rewrite but some codes from the previous versions have been used.

• Board Representation and Move Generation - It follows that of Gerbil using the 8 x 16 array for the chess board. This may be the simplest and fastest method for the PC. As explained in the codes of Gerbil, this method has an efficient way of attack detection as the "del" or distance between 2 valid squares of a move is unique for the type unlike that of using th 8 x 8 chess board. The details could be found in the codes of Gerbil. Most codes for generation and attack detection are compact and short.

  The 32 chess pieces are represented using 32 structures that contain square positions, color, type , an on-off flag and the pieces are toggled on or off as needed. This seems better than that of TSCP that keeps track of a 64 piece-board and having to loop through 64 squares instead of 32 pieces.

• Search Algorithm - The default is Nega-scout but may be set off in which case Simple Alpha Beta is used. Nega-scout may be slightly faster and seems without bug. Has the usual check extention of depth + 1. Fractional depth extention and reduction is also used. Quiesence search with only capture moves and pawn promotions. The maximum search ply is 50 and the maximum full search ply is about 30.

• Move Ordering - This is considered important by most chess programmers. Alpha-beta starts with a maximum search window of (- infinity, + infinity) and narrows as the search proceeds. When good moves that return high values are tried first, the search window narrows very rapidly and alpha-beta completes by searching less move and therefore returns faster. Lesser time needed for every search means the program can search to a higher depth for a given time.

  Presently moves that captures at least the same ranking is given the highest ordering without additions. For other moves, history heuristic is used. Also a fairly comprehensive attack scoring is added and this may reduce the overall NPS, but seems good.

• Evaluation - Evaluation for board positions may be the most difficult part of chess programming as it involves putting in chess knowledge for alpha-beta to use. A good or a bad and unreliable evaluation may differentiate between a strong and a weak program. The algorithm alpha-beta returns with the best move for the side on move and in its decision-making it relies only on one signal, the return value from calling the evaluation function; Simply, it is -
  If return value V > previous return value V-Best
  V-best = V

  So when the signals are unreliable, the recommended best move may not be too good against an opponent coming with a more reliable probability of the goodness of move. Crafty is a near grand master level chess program and its evaluation codes seem to be highly complicated with exotic terms like tropism ... and other things that need chess knowledge to understand.

  The evaluation of SnailChess follows the basic from TSCP and added other elements from common sense. The evaluation of TSCP is a raw and short core evaluator that uses the bare minimum of chess knowldege, but the most important aspect for each of the types.

  Pawns - penalty for being lone, doubled, backward. A high bonus per rank for passed-opponent-pawn and exposed, ie. cleared ahead of any opponent pawns. Bonus for advancing pawns.
  Knight and Bishop - Only piece-square table values are used, emphasis being on the central positions.
  Rook - Bonus for rook on 7th rank, on open or half-open file from blocking pawns.
  Queen - no evaluation.
  King - King penalty when it goes higher in rank before end-game. 3-pawn cover when the king has gone into a corner; Pawn storms by opponent pawns getting close to the king; Castling
  

  With the above evaluation, TSCP can out-search and out-play other programs that may be more complicated or may even have a hash table but less well implemented. So what it does not detect at raw basic evaluation, it detects after searching deeper with very fast evaluation. It simply optimizes on the trade-off between evaluation knowledge and search speed. TSCP is near the top among the basic winboard chess programs.

  SnailChess includes all of the above with minor variations. It added others like evaluations with binary-bishop and binary-rook, lone-bishop, bishop trap by pawns at 7th rank side file, knight trap, rook trap, knight passed-pawn, knight near the center of gravity of king-king and best-passed-exposed pawns of the 2 sides during end-game( maybe what knight tropism is ?), trapped rook at a corner by own king, pieces getting to the same side of a king with a weak 3-pawn cover, evaluation for the best passed-pawn, and others. Connected-pawn-chain has a high emphasis. pawn, knight and bishop attacks with double attack, and other attacking evaluations.

  If all options in the snail.ini file is set off including the hash table, then the program relies only on move ordering and evaluation.

• Hash table - Chess evaluation is difficult because of the knowledge needed to analyse the goodness of pieces on a chess board. Hash table implementation may be the next most difficult thing to implement in chess programming. The Alpha-beta algorithm is very simple and its outline may be in just 10 lines of codes; the idea of a hash table is also very simple - to store information of nodes searched into a database to be retrieved and used efficiently later, whenever possible, in place of seaching alpha-beta which is very time consuming. The difficulty is not with the programming which in itself is rather straight-forward , but the way to use it properly that is the most tricky part. Someone somewhere, maybe a certain professor of computer science, Marsland, mentioned that ... alpha-beta interacts with the hash-table ... in a strange manner.. and it seems to be true from the author's experience after doing it fresh the nth time. The most tricky part may involve hashing nodes with check-mate values and returning such values from alpha-beta.

  SnailChess uses 10 bytes per record and the hash table may be disabled or the size set from the executable command line/ init file and has a maximum limit of about 700MB. It will round up to the nearest reasonable size value. The actual value allocated from memory is printed to the log file at the start of the program. All nodes are hashed including those for quiescence search and nullmove search using the same hash-table.

• Null Move Search - This has a reduction of depth - 1 and allows for recursive nullmoves without limit. A confirmation search is done as suggested by Dieter, the author of Yace.

• Futility Prune - This is done with a threshold value of 3 pawns. Also there is an equivalent of prune for skipping evaluation and returns the material difference only when it is much greater then beta or when falling far below alpha but seaching without calling evaluation. Disabled in Ver 4.012 as it does not seem to work.

• Internal Iterative Deepening - The idea is mentioned in the codes of Crafty by Bob Hyatt but the codes are too complicated to decipher so it is done from first principle. Internal Iterative Deepening may only be needed when a hash table is implemented. Under normal alpha-beta search the principle variation is passed down from the decendent nodes back to the parents so that at the next root iteration the moves from the last principle variation may be tried first as searching out the the good moves first is important for alpha-beta to return fast. When a hash table is implemented this principle variation may be lost and not available for the next root iteration. The idea is that whenever the search is following the PV and the PV move is not available nor is there a move from the hash table, a search is done with a lower depth -1 to produce a best move to try first for the normal depth search. It seems most strong programs implement sort of this and SnailChess has it.

• Others - Certain nodes are researched with a higher depth when found exceptionally good. Fractional extension applied when nodes likely to be the PV. Some other unconventional pre-pruning.

• Opening Book - Without one at present. As the hash table is not cleared after a game, the hash table somewhat substitute as an opening book after the first game and it may play stronger after the first game.

• Log file - Some details of the game is printed to a log after every move including NPS, scores etc.

• Winboard Features supported - It supports only the basic features that enables normal playing or between 2 machines. Edit mode and playing from saved games is supported. No pondering. All time controls available.

back