This is history of Phalanx. Dates are in the form YYYYMMDD.
I First version to play a legitimate game, it finally knows all rules.
19970304 I spent a lot of time on tuning transposition table code and
extensions. The program scores 78% at large Reinfeld's test set, 10 s
per move, on 486+/150.
II Two killer heuristics added: well known history killers (modified)
199704?? and `reaction' killers (the killer is connected to the previous enemy
move). The tree search is faster, especially at `quiet' positions.
Hashtable presearch. Modified hashentry replacement strategy. Better
evaluation function, it now knows more about passed pawns, the
endgame play is much stronger. Fixed bug at 7-th row pawn push
extension routine. Faster output - using sprintf() to prepare the
output and then only one printf. New command line options: -t, -T to
set the hashtable size.
III Permanent brain (pondering). King safety evaluation. Hung pieces
19970529 static evaluation. Null move pruning. Forward pruning. Better time
heuristics - forced moves are played quickly. New cmd-line options:
-f, -x, -p, -s.
IV New option -c to determine whether to use cpu or real time. The
19970721 internal timing resolution changed to 1/100 of a second. Xboard
compatible editing position. Removed en-passant capture from
quiescence search. Changes in static-eval function. Piece list
implemented - speedup is between 0 and 20%; more in endgame. Pinned
pieces static evaluation. Better evaluation of weak pawns. Trapped
bishop and knight evaluation. Bugfix at extending check evasions.
Bugfix: development bonus was computed but not added to total
evaluation. Output change: '!' means turn, '!!' means value out of
window.
V Mostly interface changes, no big improvement in playing strength.
19970803 Xboard compatibility fixes: 'post' is no more switch, new command
'remove'. Fixed command 'level': 'level 0 5 0' (five minutes per
game) should work, also increment (ICS) levels ('level 0 2 12').
Increment can be given at command line. New option -o (polling
input) - when running tests, phalanx needs -o-. New feature of
autotesting: all incorrect results are written into file
'notfound.fin'. Small output changes for better compatibility with
xboard 3.6.2. Eliminated some bad turns in evaluation. Better king
safety evaluation - counting safe checks available.
VI Fixed small efficiency bug in evaluate.c. Improved time heuristics
19970915 for increment (ICS) levels. Optimized hashing - about 5% overall
speedup. Fixed bug in ptime(). User is now allowed to continue play
even if the position is drawn by 3-rep., material, or 50 moves rule.
Fixed bug - Draw requests are now always ignored, not always accepted
:-). Xboard-compatible 'Illegal move' message. Binary opening book,
created from PGN by bcreate. This distribution is bigger than any
previous because it contains a small example of binary book
('sbook.phalanx', 174kB, 29634 moves). SAN input accepted. Static
evaluation now better understands middlegame positions without
castling. Added pre-computed tables into static evaluation -> about
5% speed improvement. 'Show thinking' mode now shows also a list of
book moves if there's a book entry. I have played a match with
previous version (64 games, 2 minutes, increment 12 seconds, on
a 486+/150, via xboard, no pondering), new version wins 38.5-25.5.
VII Courtesy of Pavel Janik ml., new command 'fen'. Bugfix in command
19971022 line parser, 'phalanx 0 2 12' now works. Better time controls: added
hard time limit to avoid being flagged. Pondering bugfix: Phalanx
was pondering O-O, but opponent played O-O-O; the move check was
incorrect and Phalanx thought that O-O was played; this resulted in
'Illegal move' message few moves later. More pondering safety: the
move to be pondered is now checked for legality, before this fix it
was simply taken from PV (if present), but PV is not always 100%
correct. More PV safety: search cannot be interrupted during the
first iteration, the abort is delayed until the first iteration is
finished. New options: -r <resign value>, -b <+/-> to set opening
book on/off. New commands hard/easy to set pondering on/off.
Optimized do_move(): 3% speedup in test positions. Search
optimization: null move is not played if the value in hash table
entry is too low. Output bugfix: output is now readable even at long
time searches. SEE (Static Exchange Evaluator) and more pruning in
quiescence search. RoboFICS compatible commands 'white' and 'black'.
Routine for evaluating pawn endgames. New extension trick that helps
in some horizon-effect type positions, it's based on measuring
difference between current score and null move result; if current
score is much better, the side to move must be under threat and the
line is extended, similar trick is also in gnuchess.
VIII Fixed bug in search(): eliminated useless re-searches -> about 3%
19971103 speed improvement. Noise level is now in centiseconds, not in nodes.
Improved horizon-effect extensions. endgame.c: new endgame
knowledge, mostly for endgames with minor pieces and pawns: knight
and bishop mating, knight+knight (draw), minor vs. minor (draw),
minor vs. pawn (draw, but still needs some work, sometimes a pawn
wins against a knight), onepawn() as a special case of pawns(), bad
bishop (bishop + rook pawn), basics for R+P vs. R, bishops with
opposite colors (this is often drawish ending). New penalty for
moving castling-side pawns in middlegame.
IX Function smove() is now used in both binary book and hashing code, it
19971205 is cleaner and easier to read, the binary is smaller with this.
Bugfix: Phalanx now does not ponder move that leads into a terminal
position (checkmate, stalemate), this fixes some unexpected quits
(I hope all of them). Added more limits to horizon-effect extension
trigger - there was so many extensions, that version VIII was
actually weaker than previous versions, because of lower nominal
search depth; these extensions are now used only for major threats.
Static eval cache entry is now packed into 4 bytes instead of 8 ->
the cache eats only 256K (512K before this).
X Bugfix in static evaluation of trapped knight at [AH]7. Better
19980108 evaluation of trapped bishop ([AH]7) and knight ([AH][87]), now it's
using the static exchange evaluator. Time allocation: alloc more
time (+1/8) if pondering is on. Better understanding of 'trade-down'
bonus - if there are no pawns, you often need a rook more to win the
game, a minor piece is not enough. Shallower extensions resulting in
greater average search depth. Easy levels: use new command line
option -e <1...100> to select an easy level; 1 is the strongest and
100 is the easiest one, Phalanx tries to make human-like blunders.
Minor changes in resigning. Clean up in pbook.phalanx, deleted over
100 positions. Killer heuristics now uses the SEE. Bugfix in
endgame.c: trade bonus did not work for kings and pawns endgames.
Simple learning, can be activated by -l+ option; it is off by
default, because it's very experimental version. New version wins
engine-engine match with version IX 23:17 (tc 2, inc 8, 486+/150,
no learning, no pondering).
XI Courtesy of Milan Zamazal: long options and user-friendly handling of
19980125 book files - you can now use environment variables or command line
options. Bugfix in trade bonus: endgame KQ-KBPP was evaluated only
about +1.10, but the stronger side is winnig here even without pawns,
now it's >+3.00 and Phalanx can solve this Dufrense&Mieses study:
8/3Np3/7P/1p3P2/1k6/5K2/1b6/8/w Pf5f6. Deeper overall extensions,
+10% of ply. More check evasion extensions. Wider window for 'lazy
evaluation', search is now slower but stable. Minor fixes in forward
and null move pruning. Zero-width search modified for root moves,
actually it's no more zero-width; this eliminates some 'blind turns'
(a move is evaluated as >=(Alpha+1), but the engine cannot prove it
in re-search in full [Alpha,Beta] window). Minor changes in static
evaluation. Bugfix in time allocation of increment levels. Changed
time heuristics to use more time.
XII More pruning in quiescence search: moves that have little chance to
19980301 return the value over alpha are not generated. Permanent brain is
now really 'permanent', if there is no move to ponder in the PV,
Phalanx uses short search to pick a move and ponders it. Bugfix:
occasional zero division at very fast time controls, like 10s/game.
Some puts() replaced with printf(), puts() is buggy in my gcc (2.7.2)
and causes sigsegv when interrupted and called again during the
interrupt. Optimized null move, this speeds up the engine. Easy
levels now use nodes instead of centiseconds for measuring time, the
playing strength of easy levels does not depend on cpu power and
cpu load. Resigning now works with RoboFics (via 'tellics resign').
Updated FSF address in docs.
XIII Lots of 'char' and 'short' replaced with 'int' -> speedup on iX86
19980316 chips (about 10% on my 486, probably more on P6). Bugfix at 'edit'
command, edit was broken for positions with black to move. New
command line option 'bench', 'phalanx bench' runs a 5 cpu-minutes
benchmark and reports nodes per second. Added some interesting lines
into the opening book. Dynamic draw score, it's set to -20 by
default. New command 'rating <phalanx rating> <opponent rating>' to
adjust the dynamic draw score, Phalanx tries to avoid draw with
weaker opponents, minimal draw score is -20, maximal +20.
XIV Dynamic draw score did not work for stalemate, now fixed. Major bug
19980416 fixed at resigning: if resigning was on, phalanx resigned almost any
game after few moves, this was broken from previous version.
Simplified computing bounds for lazy evaluation, this was too
complicated and slow. Higher material limit for evaluating position
as endgame. Deeper extensions, +0.1 ply. Improved move ordering at
root search. Minor optimizations. More variability in opening book.
Disabled scout search near leafs. Bugfix: type of Nodes changed to
'long long' (64 bits on Intel[3456]86), 32-bit integer was not long
enough for >24 hour runs. Bugfix: option -P did not work, now fixed.
Added startup messages that show full path of all open files.
XV Internal iterative deepening. Cleaned pbook.phalanx. New "Pc2-c4"
19980614 code to encourage playing c4 (c5 as black) in closed positions where
e4 (e5) is not possible. Bugfix at computing total material on the
board, the variable was sometimes overwritten with unexact value.
New xboard compatible commands 'bk' and 'analyze'. 'bk' shows all
book moves. 'analyze' starts analysis mode. Bugfix in bcreate.c: no
games on stdin caused segmentation fault. GNU extensions (long
options and snprintf) are now used only if GNUFUN is defined
(-DGNUFUN in makefile). New endgame knowledge: KNP vs. K is
sometimes draw. Small changes in static evaluation. SIGINT is
ignored in polling input mode. Time information in post lines is
shown in centiseconds instead of seconds to follow current draft of
xboard chess engine communication protocol.
XVI Bugfix in analyze mode: there is no extra character after the search
19980630 depth number in post lines - xboard understands that time info is in
centiseconds, not in seconds. Bugfix in null move pruning: the side
that is trying its second move in row must have at least one legal
move, this caused segmentation fault if there was none. Command
'hard' does not set pondering on in easy levels. Bugfix at parsing
SAN moves, nc3 was (possibly) interpreted as Pc2-c3. Command 'level'
accepts time in mm:ss form, for example 'level 0 2:30 12'. Bugfix in
kings+pawns endgames - penalty for pawns on one file (doubled) worked
as penalty for pawns on one row. Improved kings+pawns endgames
static evaluation. New command 'xboard' sets xboard compatible mode.
XVII More variability in pbook.phalanx. Better sbook.phalanx, generated
19980804 from wall.pgn. Changed (and fixed) extensions for check evasions and
passed pawn pushes. Changed evaluation of passed pawns. phalanx and
bcreate integrated into one binary, most of the code was common
anyway; book can be created via 'phalanx bcreate' instead of just
'bcreate'. Bigger limit for count of [position,move] in bcreate
code, the counter has 32 bits instead of 16. Performance bugfix in
null move pruning.
XVIII Bugfix in the static exchange evaluator - it let kings slide
19980906 horizontally. New styling of board in command "bd"/"d". New command
in edit mode: 's' to switch side to move. Modified look of edit
mode. Modifications in static evaluation: bishop pair needs
mobility; queen has 'king distance' bonus only if it's on a safe
square; simpler rook evaluations - smaller bonuses; forpost (outpost)
pawn gets more bonus for # of pieces defending it (Aaron Nimzowitsch,
"Mein System"); bigger bonus for protected or connected passed pawns.
Modified threat extensions - only mate threats are extended. Bugfix
in analysis mode: PV is computed (and shown) even if there is only
one legal move. New feature of command 'level' - 'level N' sets
level to N seconds fixed time per move. Better pbook.phalanx.
XIX Log file to allow easier debugging/tuning, it is activated via
19981122 '-g <log file name>' command line parameter. Positional learning
file is created (in current directory) if it does not exist and it no
more has to be distributed as empty. Minor speedups in pawn
endgames. Outside passed pawn detection and evaluation in pawn
endgames. Raised material value needed for evaluating
kingside/queenside storming (attack after different castlings), these
are not worth evaluating when close to endgame. Fixed book move
selection to make it easier to implement book learning in future (no
real change for user). FICS whispering. Better pbook. Bugfix in
evaluation of kings+pawns endgames - the evaluation of 'king in the
square of pawn' had bad sign (!). New extension trick: we add big
extension after two succesive non-pawn captures that result in pawn
endgames - the branching factor is much smaller below that node and
we can go deeper without affecting search stability and ... we should
do so, because 6 plies subtree evaluation in middlegame is much
better than 6 plies in king+pawns. Timing change: phalanx uses more
time if there were turns in last ply level. Fixed bug in
communication with xboard that caused playing too fast in '-mg' mode
with 'reuse' on. Alias 'u' for 'undo'.
XX Bugfix in xboard mode, long lines of 'tellics whisper ....' were
19990104 split into several lines. Better pbook.phalanx. Fixed code that
avoids pseudocastling (moving white king e1-f1-g1 and leaving rook h1
undeveloped), now it works also in endgame. New code to keep
pressure against stonewall d-pawn; if black has stonewall, white
should keep the c-pawn on c4; previous versions often played c4-c5?,
and that is a bad error against humans. Knights now get penalty in
endgames with pawns on both sides of the board (12 to 35 centipawn).
Outside passed pawn evaluated in all endgames, not just in
kings+pawns. Minor bug fix in analyze mode with xboard (phalanx
aborted search due to unimplemented '.' command). New side effect of
command 'level' - it also sets remaining time to the same value as
level time; this does not break anything but might help in the
Fritz 5.32 winboard interface. SAN input now accepts pawn captures
in the form 'ef'. ECO database, ECO codes and names are shown with
'bk' command; user has to run 'bk' command once in initial position
to activate the ECO feature. Removed '-T' command line option from
help, it might be used to tablebase directory in future.
XXI Major bug fixed in outside passed pawn evaluation, when one side had
19990131 no pawns, there was indexing out of range. Minor penalty for rook on
a pawn file in middlegame. Fixes in pbook.phalanx, as usual.
XXII Better ICS whispering, it sends only one info per move. Bugfix in
20000116 disabling castling flag, now also moves _to_ the A1,H1,A8,H8 squares
disable corresponding castling. Improved pbook.phalanx. New
sbook.phalanx created from "granit" collection. Modified
bcreate so that moves are put into the book in the order of their
success rate, the most successful move is later shown as the first
one in the list of the book moves. Bcreate now uses temporary files
instead of memory. Deeper extensions. Modified and fixed rook
static evaluation. Learning is now on by default. Phalanx sets
the first finger line on ICS, it shows version, size of hashtable,
and sizes of the opening books. Bugfix in benchmark, the resulting
number was incorrect on machines that can do over 70kNps.
