Your trusted source for computer chess information!
First of all, thanks very much to Ron Murawski for making this site available and hosting Myrddin here!
Version 0.87 to be released soon! This version supports SMP via processes (up to 16 instances) and is about 10 ELO stronger than v0.86 at 1 CPU and about 50 ELO stronger at 4 CPUs. This version has not been formally tested above 4 CPUs so it is unknown how well it will scale above 4 CPUs.
Myrddin 0.86 (released 12/21/12, CCRL rated approximately 2350)
This package contains Windows32/64, Linux32/64, & Android executables compiled by Jim Ablett, and a Mac executable compiled by Julien Marcel, author of the Prédateur chess engine.
Myrddin 0.85 (released 5/3/11, CCRL rated approximately 2215)
Myrddin 0.84 (released 9/17/10, CCRL rated approximately 2095)
Myrddin 0.83 (released 2/22/10, CCRL rated approximately 2000)
Myrddin 0.82 (released 9/26/09, estimated CCRL rating of 1915)
Myrddin Alpha 2 (released 5/26/09, estimated CCRL rating of 1550)
Myrddin Alpha 1 (released 3/9/09, estimated CCRL rating of 1150)
Myrddin 0.87 – 1/20/15
Myrddin is a winboard-compliant chess engine to a reasonable degree. It supports protover 1 primarily to support the Chessmaster interface, and also protover 2.
If you don't know what Winboard is, and you only just want to play against this engine, you can download Winboard-compliant interfaces here:
- Winboard (the original, but not easy to set up)
- Arena (very popular and supports both Winboard and UCI communication protocols)
- Aart Bik's excellent Chess for Android program is recommended for trying out Myrddin's Android version.
The 64-bit version of Myrddin should play approximately 2380 ELO against chess engines at 1 CPU (using CCRL rating), and probably 100 points higher against humans. This is an improvement of approximately 30 points compared to the previous release. Myrddin should play about 90 points stronger at 4 CPU.
The following winboard commands are supported:
- setboard (also “loadfen”)
Myrddin also supports the following non-winboard commands:
- “eval” – returns a static evaluation of the current game position
- “perft N” – standard perft algorithm to calculate the number of leaf nodes of depth N from the current position. Note that Myrddin will not respond to any input until the perft calculation is finished.
- “divide N” – extension to perft algorithm showing the number of leaf nodes of depth N from the current position, divided among each legal move from the current position.
- “tb” – toggles Gaviota endgame tablebase support
None of these commands are supported while Myrddin is searching/analyzing.
– It is crucial that the winboard UI send the “time” command to the engine, as Myrddin does not have an internal clock.
– Post is ON by default, as opposed to the winboard protocol. This is just for debugging convenience and it appears that a lot of engines do it this way anyway.
– This is the first version to support SMP. The implementation is based on Dan Homan's “lazy SMP” idea. Myrddin uses multiple slave processes to fill the transposition, eval and pawn hash tables so the parent process can search deeper in the same amount of time.
– Myrddin uses Pradyumna Kannan's magicmoves code for move generation of sliding pieces.
– Evaluation is rather simple: wood counting, piece square tables, pawn structure, rudimentary king safety, rooks behind passers. There's still a long way to go here.
– Search is basic alpha/beta, with reasonable and generally conservative extensions and reductions.
– All user-modifiable parameters are supported in the INI file, an example of which is included in the download.
– Max search depth is 128.
– The ProDeo opening book is used by kind permission of Ed Schröder.
– Draw claims from the opponent are not supported. Myrddin does, however, claim all draws by rule, as well as checkmate.
– There is enough winboard support to play games on ICS. But without support for “draw” I'm sure there are some scary loopholes and/or exploits.
– When the engine is in analysis mode, positive scores always favor White and negative scores always favor Black. When the engine is thinking or pondering, positive scores favor Myrddin.
– Logfiles will be in the “logs” folder below the folder where you ran Myrddin. The output of the log is not very interesting – just PV output and communication reality-check stuff. If you are running Myrddin with multiple CPUs, there will be one logfile for each process.
Myrddin's first official tournament was in the Promotional division of ChessWar XIV and was rated at 1655 over 11 rounds. Myrddin scored 6.5 points (+5 =3 -3) and placed 51st out of 212 engines rated between 1859 and 812. The median engine rating was 1499. To see the PGN of Myrddin's games and see some comments, read this journal. Much of the development between Alpha 1 and Alpha 2 occurred during this tournament.
Myrddin participated in OpenWar 6. This was a round-robin tournament with 92 engines. Myrddin finished in 68th place with 27 points (+22 =10 -59). Myrddin's best “legitimate” results were a win against Firefly (rated about 2050) and a draw against Timea (rated about 2150). It managed to get a win against one of the top 10 (Cippolino) thanks to a crash. It also got two draws against two very strong engines (DanaSah and N2) due to endgame blunders, and a win against another top engine (Tornado) due to a crash. Much of the development for version 0.82 occurred during this tournament, and the last few rounds were played with 0.82.
Version 0.82f (unreleased and slightly better than 0.82) participated in the 5th Division of WBEC Ridderkerk 17. There were 84 engines in this division, and the division was broken down into three groups of 28 engines each, each group having a double round-robin tournament. As expected, Myrddin finished in the middle of the pack, in 14th place with 31.5 points out of 54 games (+28 =7 -19).
Version 0.82f also participated in the Promotional division of ChessWar XV. This tournament was very similar to Myrddin's first tournament, ChessWar XIV, except there are even more engines (242!). Because Myrddin was at least 300 points stronger than it was eight months previous, it finished in 9th place with 8.0 points (+7 =2 -2). This was a good enough performance to promote to the next division, Myrddin's first promotion. Myrddin now has a bit of a reputation as a very lucky engine, as it got a free point when GnuChess (a typically very stable engine) crashed in a Mate in 3 position. Without that crash, Myrddin would have finished about 20 places lower and would not have promoted.
Versions 0.83 through 0.83k participated in OpenWar 7. This was a round-robin tournament with 76 engines. I was hoping that Myrddin would place around 55th and score about 25 points. But, despite some noticeable improvements in Myrddin that went into version 0.84, Myrddin performed poorly and finished in 62nd place, with only 20.5 points (+14 =13 -48). Myrddin's best results were draws against the 11th place engine, WaDuuttie, and the 20th place engine, Rotor. Both of these engines are rated 2500+. Its best win was against the 45th place engine, Sungorus, rated 2300+. The only other good thing that can be said is that Myrddin had good results against the 18 engines that finished immediately above him (+3 =9 -6).
Version 0.84 participated in the F division of ChessWar XVI. There were 80 engines in the tournament and Myrddin was seeded 70th. Myrddin finished respectably in 49th place with 5 points (+3 =4 -4). Its best performance was a draw against the engine that finished in 6th place.
Version 0.84 (32-bit) was the first version to win any tournament, doing so in the 23rd Series the CCRL Amateur Division 7. It competed with nine other engines in a double-round robin and won with a score of 23.5 out of 36 (+20 =7 -9). This victory promoted Myrddin to Division 6.
Version 0.84g (unreleased and clearly better than 0.84) participated in the 5th Division of WBEC Ridderkerk 18. There were 98 engines in this division, and the division was broken down into three groups of 32-34 engines each, each group having a double round-robin tournament. The hope was that Myrddin could place in the top 6 in his group, thereby reaching the final. Myrddin did accomplish this, scoring 55 points in 66 games and placing 4th. The final was another double round-robin with the top 18 engines (6 from each group), and the top 7 would promote to the 4th division. Myrddin managed to achieve this by placing 7th, scoring 18 points in 34 games.
Versions 0.84g through 0.85 participated in OpenWar 8. This was a round-robin tournament with 78 engines. I was hoping that Myrddin would place close to 40th and score about 37-38 points. This tournament was a mixed bag, as Myrddin drew 10 of the engines that finished between 15th and 30th places (and all 6 engines between 15-20: Brutus, Hermann, Dirty, Arasan, Francesca MAD and Gaviota). But some poor performances against weaker engines kept Myrddin from achieving my goal. Myrddin finished in 46th place with 34.5 points.
All results are for the 64-bit version of 0.85 using 128MB hash and 5-man Gaviota tablebases. The tests are run on an AMD 9650 Quad 2.3GHz machine (although Myrddin does not support multiple CPUs). Running all of the below tests takes approximately 5 hours.
300 positions at 5s per position. Almost exclusively tactical positions, many leading to mate. Myrddin finds 295 correct moves within the time limit. Regarding the five missed positions, two (141, 222) are found within 10 seconds, one (293) is found within 30 seconds, and two (17, 230) are not found within one minute. This suite can be found here.
A series of (currently, but occasionally expanding) 14 individual suites of 100 positions each, primarily intending to test positional knowledge such as square vacancy, proper recapturing, pawn advancement, control of the center, etc. At 10s per move, Myrddin gets 789 out of 1400 positions correct. This suite is also interesting because it contains “partial credit”. In other words, you can still get some points by choosing moves other than the optimal. Myrddin gets 9,556 out of 14,000 points if partial credit is included. You can download this test suite here.
176 positions at 10s per position. Combinative middlegame and endgame suite. Myrddin finds 134 positions within the time limit. You can download this test suite here.
100 positions at 10s per position. Myrddin does not do very well at this suite, getting only 20 correct moves within the time limit. You can download this test suite here.
182 positions at 10s per position. Challenging tactical suite. Myrddin finds 149 positions within the time limit. You can download this test suite here.
For a list of (hopefully) interesting test positions that I have accumulated during the development of Myrddin, go here.
Many parts of the code from Myrddin contain ideas and/or algorithms gathered from studying the source code of other engines. Most notably, Stockfish, Fruit, Crafty and Greko are well-written, well-documented and relatively easy to understand even for a mediocre programmer like myself. The list of features implemented (and sometimes modified or discarded altogether) from these ideas would be too long to go into here. But I can safely say that all code in Myrddin is my own, written either from scratch or based on these ideas, with the following exceptions:
– Myrddin's Winboard interface is based on Tom Kerrigan's excellent TSCP engine, for which Tom has graciously given permission.
– Myrddin's SEE implementation is based on pseudo-code generously given to me by Andres Valverde, author of EveAnn and part of the Dirty development team.
Initial Release, rated approximately 1200-1300
– Reduced move generation time by ~21% (perft 6 on initial position went from 57 to 44 seconds on P4-3.0)
– Search now pings the input command handler every 8K nodes (about 1/30-1/50 second, depending on position and hardware), so applicable commands can now be entered during search/analysis
– Search will allocate extra thinking time when it gets noticeable drop in score
– Evaluation improvements – pawn structure, open and semi-open files for rooks, rooks behind passers, king safety
– Added MVV/LVA and PV Move Ordering
– Added Killer and History move ordering heuristics
– Added Null Move reductions
– Added Late Move reductions
– Added Opening Book (ProDeo – Thanks, Ed!)
– Added pondering, so “hard” and “easy” commands are now supported
– “analyze”, ”?” and “result” commands are now supported
– “level” command is now fully supported for all time control types
– Added check for dirty pawn structure before evaluation
– Search can be interrupted due to time management considerations
– Added Hash Tables, with a fixed size of 128MB
– Maximum search depth increased from 20 to 30
– Fixed a bug with evaluating castling moves
– Fixed a bug determining which King piece table to use
– Can now generate capturing moves only so quiescent search has less moves to sort/deal with
– Quiescent moves are added to PV
– Added code for pushing lone King towards edge of board
– Added material and 50-move draw detection
– Fixed a bug in the “undo” and “remove” commands
– Fixed some bugs in the 3-fold repetition detection
– Fixed a stupid bug that caused me to check for user input WAY too often – many thanks to Bob Hyatt for pointing me in the right direction
– Significant code cleanup
– Removed “Alpha” designation from version number
– Commandline parameters are now supported
– Increased max search depth to 50
– Reduced move generation time by ~15% (perft 6 of initial position on a P4-3.0 went from 44s to 38s)
– Now claiming checkmates, 50-move draws and 3-fold repetition draws
– Plays a move after completing a depth 3 search if there is only one legal reply
– Will not play a move if in the middle of resolving a fail low or fail high at the root
– Will never use more than 1/4 of remaining time
– Greatly improved time management for bullet games. Version 0.81 could lose as many as 20% of its games on time at 2 minutes per game. Version 0.82 now only loses about 1% of its games on time at 1 minute per game, and only very rarely will lose a game on time at other time controls. Many thanks to Lars Hallerstrom (The Mad Tester!) for all of his assistance in helping determine if my “fixes” actually improved anything. Myrddin should never lose on time in games with increments.
– Quiescence search now only searches recaptures after depth 1
– Added promotions to Quiescence search
– Added aspiration window
– Improved Late Move Reductions parameters
– Fixed a bug in the Principal Variation search
– Pondering is now OFF by default
– Evaluation adjustments:
- bishops and knights are scored at 310 centipawns, as Myrddin was susceptible to trading two pieces for rook+pawn
- increased penalty for doubled/tripled pawns
- for doubled/tripled pawns, added further penalty if they are blocked
- added larger bonus for passed pawn on the 7th
- adjusted bishop piece tables to encourage occupation of long(er) diagonals
- modified the lone king piece table to give larger penalites as the king goes towards the edge/corner
– Added knowledge of knight outposts (in enemy territory, supported by pawn, cannot be attacked by enemy pawn)
– Piece and wood counting are now done incrementally
– Fixed a bug in which the 50-move counter could be set to zero if the move being pondered was a zeroing move and the opponent did not play that move
– Fixed a bug in which the engine would go into an endless loop if it reached (max_search_depth + 1) during pondering or analysis
– Fixed a bug in which the engine would go into an endless loop if there were no legal moves during analysis
– Fixed a bug in the evaluation of Black doubled/tripled pawns
– Added code to handle KPvK endings
– Added code to recognize various material draws, such as KNNvK
– Added code to recognize insufficient mating material for materially winning side (e.g. KNNvKP or KBvKP)
– Added code for rook pawn and wrong colored bishop against lone king in promoting corner (from a loss to Sorgenkind in OpenWar)
– Now only adding en passant square to hash signature if en passant capture is possible
– Added ”?” and ”!” comments to PV output for fail lows and fail highs
– Logfiles are now created in a “logs” folder below the Myrddin executable program, as requested by Lars
– Only search to depth 1 when there is only one legal move and pondering is off
– No longer clearing the hash table before starting a depth 1 search
– Improved hash replacement strategy
– If a search depth was reduced by late-move reductions, and that search improves alpha, now researching at proper depth
– Reductions are now less aggressively implemented
– Adjusted the second aspiration window from 300 to 110 centipawns
– Doubled (or worse) passed pawns after the leading pawn no longer get the passed pawn bonus
– Will now play a move in a checkmating line as soon as it is confirmed to be optimal
– Removed “Result” from checkmate reporting
– Rook and minor vs rook (with no pawns) is now hard-coded as a draw – should be safe in 99.99% of cases
– Re-removed bad king safety code
– Now using piece square table for knight outposts, rather than a hard-coded bonus
– Modified piece square table for kings in opening and middlegame
– Added penalty for having no pawns, to help avoid trading down into pawnless endgames (e.g. rook vs. minor)
– Sending the '.' command during analysis no longer restarts the analysis from depth 1. The '.' command is still not implemented, though.
– Fixed a bug in which most hash moves were not getting placed at the beginning of the move order
– Fixed a bug evaluating Black doubled and passed pawns
– Fixed a bug in which the initial position was not being checked for draw by repetition
– Fixed a bug maintaining wood tables when unmaking a promotion move
– Fixed an asymmetrical problem in the bishop piece square table
– Fixed a problem with determining if a rook or queen was behind a passed pawn
– Fixed a problem with calculating the pawn shield of a king on column h, or a king not on his home row
– fixed a bug updating the board signature after making a null move when en passant was possible
– king+rook vs. king+minor is now being scored as a draw
– Added Gaviota tablebases (thanks so much to Miguel Ballicora, also the author of the Gaviota chess engine, for making this available!)
– Added initialization file for setting hash size, creating log file, turning on kibitz, and tablebase info
– Fixed a bug in which a hash probe would not return a hash move if the saved depth was less than the requested depth (thanks to Edmund Moshammer, author of Glass!), so in these cases the hash move would not be used for move ordering
– Implemented fail-hard
– Tweaked parameters for reductions (yet again)
– Adjusted queen value from 900 to 950 and minor piece value from 310 to 320
– Adjusted bishop pair bonus from 20 to 30
– Removed second aspiration window
– Lazy eval is now more conservative
– Will now analyze the entire depth if the first move searched fails low, or if first move fails high in a non-mate situation
– Increased maximum search time to half of the remaining clock (required due to above item)
– Added support for winboard “computer” command – kibitz PV, book moves and “Mate in N”
– Increased max half-moves in a game to 1024
– Fixed a bug in which the first move searched at the root was saved in the hash table as an exact value even if it did not improve alpha
– Fixed a bug in which Myrddin would run out of time by thinking indefinitely if it was told it had a negative amount of time on its clock (can happen on ICS due to lag)
– Modified the check to see if null move is allowable, previously was if any piece was on board, now only if side to move has at least one piece
– Almost complete rewrite of the evaluation, in particular the pawn structure evaluation and king safety. Also added mobility factor and significantly adjusted many piece table values
– Added SEE (thanks, Andrés!)
– Added futility pruning
– Added resign. See Myrddin's INI file for instructions, as this is off by default
– Added second aspiration window at 150 centipawns
– Cleaned up code for how reductions/extensions are used to modify depth in search calls
– Will now claim draw by insufficient material, but only checking for bare kings or at most one minor on board
– Now generating all moves in quiescent search if side to move is in check
– Now searching all moves in order of their move score order, instead of just the first four moves and then taking the rest as they were generated
– Some changes to the transposition table replacement strategy
– Fixed a bug in which the en passant square was not being passed to the Gaviota tablebase probe
– Fixed a bug in which the 50-move draw check was not checking that the drawing move was also checkmate
– Fixed a bug in which the flag for a mate threat found during null move was not being stored in the hash table
– Fixed a bug in which a hash entry could be saved when the engine was forcibly bailing out of a search
– Fixed a bug that could cause Gaviota tablebases folder to not be read properly from the INI file
– Finished conversion to bitboards from 0×88
– Added pawn hash (thanks again, Dann!)
– Removed lazy eval
– Added support for winboard “st” and “sd” commands
– Lots of bug fixes either discovered by the bitboard conversion or removed by them (of course, new bugs added!)
– Added SMP support for up to 16 instances (using processes and shared hash memory – Very Lazy SMP!) – only tested up to 4 instances! This use of shared memory for all hash tables (transposition, eval and pawn) means that the reported memory usage may appear to be incorrect depending on what program you use to get the information. Beware!
– Fixed very embarrassing bug in passer eval calculating the distance between two squares that could cause any number of other issues, even a hang or crash
– Added support for the “cores” command which changes the number of processes being used when SMP is enabled. The initial value will be based on ini file settings.
– Added a “tb” command which toggles tablebase support. The initial value will be based on ini file settings.
– Added “Mate in N” announcements to PV output when applicable
– Fixed a bug in saving scores near mate to the hash table
– Fixed a bug in determining valid knight outpost squares
– Fixed text output bug in kibitzing opening book moves on ICC
– Fixed a rare bug in determining material draw
– Fixed a bug in which Myrddin would hang or crash if the '.' command (used by Chessmaster GUI) was sent while Myrddin was pondering
– Fixed a bug that could cause hash memory allocation to fail if more than 1GB was requested. Thanks to Graham Banks for reporting this issue.
Many thanks to the following brilliant people who helped/guided me (either directly or indirectly) in countless ways with their work:
- Ron Murawski – Horizon and LOTS of helpful comments and suggestions, plus guidance in my first trips in 64-bit land
- Lars Hallerström - The Mad Swedish Tester! I would never release anything without his (grudging) approval.
- Dann Corbit – Without whose generous help, huge speed improvements, bug fixes and general suggestions, the first bitboard version (v0.86) would not have been released
- Dr. Robert Hyatt – Crafty (it is impossible to create an engine without thanking him)
- Miguel Ballicora - Gaviota and making his tablebase generating/probing VERY easy to implement
- Mark Lefler and the Chess Programming Wiki team.
- Ed Schröder and Jeroen Noomen – ProDeo Opening Book
- Jim Ablett – creating optimized executables
- Tom Kerrigan – TSCP
- Bruce Moreland – Gerbil
- Chris Tatham – Numpty and testing assistance (and great analysis of testing results)
- Vladimir Medvedev – Greko
- Pham Hong Nguyen – Firstchess
- And finally, the Chessmaster Development Team, comprised of lots of brilliant people, but mostly Johan de Koning (programmer of The King engine), Don Laabs (programmer of many of Chessmaster's low-level chess UI routines), James Stoddard (programmer of Chessmaster Live, the winboard UI support, and most of Chessmaster's game-handling routines) and Dave Cobb (programmer of Chessmaster's opening book editor and many other support features).
Please send bug reports and general suggestions/comments to JohnVMerlino@gmail.com.
Thanks for playing!