V3.22 searches about 20% faster and therefore has some ELO points higher strength.

I have tested this version quite much and it should run stable.

- V3.20 searches less nodes (verified null move pruning) and plays stronger as version V3.19.

- In the calculation of the Hash table index, only the low 32 bit of the hash keys are used and the upper 32 bit for collision testing.
This should lead to less type-1 errors at hash reads.

- further measures against stops of RDChess because of "wrong moves received over the WinBoard interface"  (no reset of force state at new cmd)
- In Analyze Mode: Send thinking lines (command "post") " is turned on.

- The WinBoard interface log is implemented newly and differently. Instead of logging into a main memory log, which is saved at the end of a game to disk file "WinBrdLog.txt"  a log line is directly written to the hard disk file "WinBoardLog.txt".
- V3.18 supports the Portable Game Notation (.pgn) tag "PlyCount"

- The Feature "Auto232-Play over serial I/0" is no longer supported by RDChess and has been removed from the RDChess source code.

- Message "Error at reopening OpLib. IOERROR:2" when starting RDChess with a current directory different from program directory and not using the opening book.
 - WinBoard module: AbortSuche only when MainForm.iState = iStart

- During a running search when calling the mov list windows a couple of error  messages were sometimes displayed. These are suppressed now  (procedures IsCaptureMove, IsNonCaptureMove)  
Improvements

- The move select algorithm in the search has been improved substantially  with pawn promotion moves (much better sort in move win values)
- New additional capture move generating routine (SchlagNachGenerator)  generates capture moves the other way round as the old routine. The old routine scanned all pieces from the side to move and tested  if these pieces can capture. The new routine scans all pieces from  the side not to move and tests if these pieces can be captured. This  routine is advantageous in some constellations.
- The capture move list capacity is raised to 144 (test position from Hermann  with 147 moves is handled now by default)
- Report by Leo Dijksman 7  5/40 games were lost by time! -> further reduction of  time parameters in Level Tou

- After a mouse click on the board the "from square" is not accepted if the  figure to move is absolutely pinned (TestPinned(FldFocused,PC) = UnbPin).      

- Leo Dijksman (NL) reported that versions V3.12 and V3.13 were sometimes hanging after the end of winboard games (the RDChess Windows task remaining in  memory).

 I tested V3.14 intensive  for one and a half day and the error was not reproducable, so I hope it is gone! 

Error corrections:
- RDChess V3.11 played still rarely in WinBoard Mode an illegal move when starting a new game with Black. V3.12 has now been running for 3 weeks without this error happening.

- I again changed the way RDChess follows a mating variation after finding a "mate in n moves". Sometimes RDChess came into a cycle, repeating moves and scoring a draw by move repetition instead of a clear win by mating in a few moves. It's better in V3.12, but I have seen a few games where the mating sequence was still far away of an optimum!

- When receiving a hint command from WinBoard, RDChess was sometimes responding with a wrong move (with HV[0,1] instead of HV[0,0])

- A severe error in the King-Bishop-Knight vs lone King endgame routine was corrected.
- I corrected an error in the procedure "SetWBRestZeit"  (millisecs were not counted for).

- When aborting pondering during the move of the WinBoard opponent the maximum wait for responding has been raised from 2  to 4 seconds.
 - WinBoard interface: At receiving WB commands "black" and "white"  no action is taken (commands are obsolete!).

 

- RDChess plays without problems at the American and German Internet Chess Servers FICS and GICS. RDChess displays now on its GUI in its "Game Data" window the ICS host name and the ratings of RDChess and its opponent, received over WinBoard from the chess server.
The white and black rating are stored in saved games in the .pgn tags [WhiteELO] and [BlackELO], the host name in the game annotation rich text field.
Computer opponents are marked with an appendix "(C)" in the opponents name.  

- New Popup context menu items "Accept draw from WB opponent unconditionally"and "Offer draw to WB Opponent" may be checked by the RDChess operator
-  Decoding WinBoard input: For faster parsing the first letter of the commands is checked and fast Goto Labels were inserted.
- Time management: V3.12 will use more time compared to its predecessors when searching a move. Older versions did not use the allocated time efficiently.
- Search: if Time Advantage Opponent  > 20 seconds (opponent has more time left), RDhess searches longer
-  Switch "Danger for TimeLoss" is not set in Tournament level when Move count-Difference  to Time control is < 2 (RDChess lost a game because of a too short "D"-search at the 39th move with time control n minutes/ 40 moves !!!!)
- Iterative deepening with timed searches: The constant SaveFakt was raised.
- Quiesc. search: if (max depth - actual depth) > 7 (instead of 5) then breakout of quiesc. search.
- Evaluation:
  . The Pawn/King hash table size was increased from 32 k to 128 k entries because of "mem low" with larger time controls
  . Newly introduced reward for a "knight outpost"
  . Backward pawns: changed eval. (if at squares -23,-24 no supporting pawns, then no additional punishment (seems to be quite advantageous!)
  . Pawn advance table : a minor unsymmetry was corrected.
  . Switch DiffColBishopOnly: When only single bishops of opposite colors and  pawns are on the board: -> evaluation is drawn against a draw
  . In the material advantage function "TotMaterialAdv" a maximum value of 2.40 pawn units was set for the exchange bonus.
    + Changed officer values  for queen (-3), rook(-3), bishop (+1), knight(-1). BFenster 90->80 -> 17 Hits with 3,281.706 nodes searched (-6 %) in Bratko Kopec F7 test.
 - Opening book:
 . The opening book has been enlarged. 
 . When RDChess is on move and there is no "active" move left in a board position stored in the opening book (only "passive" moves) RDChess leaves the opening book (until V3.11 it used here also a stored passive move) and searches a move by its own.
. RDChess searches with ponder on also when inside opening book! When opening book learining is on, this running search is stopped (this has become nessecary because recently a "ponder on" search on the opponents move was introduced for the case that the opponent chooses not a book move)
 . 2 new check boxes in opening book form: "learn active moves" (toggle  active/passive) and "Learn silently" (without dialogbox for inputting  position text)

Attention:

a) V3.1 has a newly introduced error at WinBoard games, sometimes claiming a wrong result  (win or loss) at the end of a game!

b) Very rarely  (reading from the transposition hash table a wrong position) V3.1 makes a serious blunder (wrong move) because I removed a test procedure which fixes such cases.

V3.1 plays stronger than V3.09.  E.g. Result of  35 games test match V3.1 - V3.09 with 4 min for 40 moves: 22:3:10  

 

 - Incremental 64 bit Hash key is calculated with xor instead of Inc/ Dec.   -> Error collisions (same hash key for different positions), which were quite  frequently in earlier RDChess versions (I never knew why!) are much less!  So few collisions, that I even do not call the collision test procedure any more.
 - V3.09 and earlier versions sometimes did'nt mate straight forward because the search  terminated when a mate in x (independent of the size of x) was found. This behaviour was especially relevant with the "Permanent brain switch" (Ponder) off and led in a few cases to repetitions and a draw instead of a win of  RDChess.
 - Further tuning of search parameters: Bratko-Kopec test suite: 3,167.023 nodes searched with fixed depth = 7 with 16 of 24 test positions solved.
 - In V3.09 and earlier versions positions with a best move at the search horizon depth (max full width depth) were hashed. In V3.1 only true full width search positions are hashed into the transposition hash table.

- a minor unsymmetry in the evaluation function between black and white pawn evaluation has been corrected.