Newsgroups: vucsm.announce,rec.games.abstract,comp.ai.games
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!cis.ohio-state.edu!math.ohio-state.edu!howland.reston.ans.net!news.nic.surfnet.nl!sun4nl!cs.vu.nl!victor
From: victor@cs.vu.nl (Victor Allis)
Subject: World Championship Awari (Results)
Nntp-Posting-Host: kits.cs.vu.nl
Sender: news@cs.vu.nl
Organization: Fac. Wiskunde & Informatica, VU, Amsterdam
Date: Sun, 1 Oct 1995 14:33:04 GMT
Message-ID: <DFrxs2.KFs.0.-s@cs.vu.nl>
Lines: 156

The World Championship Computer Awari

On September 30th, October 1 and October 2, the World Championship Computer
Awari is being held at the Vrije Universiteit Amsterdam in the Netherlands.

Three programs, all previous competitors at the 2nd, 3rd and 4th Computer
Olympiads (1990-1993) have entered the competition:

- Lithidion (Maarten van der Meulen and Victor Allis, Netherlands).
- MyProgram (Eric van Riet Paap, Netherlands)
- Marvin    (Thomas Lincke, Switzerland)

The tournament arose from a challenge by Eric van Riet Paap, who thought it
was about time that Lithidion, three time gold medal winner at the Computer
Olympiads, would allow MyProgram a change to take over.

All three programs have been built using the same concepts:
- A fast a-b search (between 20 and 30 ply per move)
- A large end-game database (between 17 and 22 stones)
- A large opening book (between thousands and close to a million opening lines).
Futhermore, Lithidion uses pn-search as a 'secret weapon' to refute mistakes by
a-b search, or to find quick wins near the end of the game.

Each of the programs has its own strong points: MyProgram is the fastest
searcher (normally 1 or 2 plies faster than Lithidion and Marvin). Marvin
has the largest opening database, while the opening databases of Lithidion and
MyProgram seem comparable. Marvin also has the largest end-game database 
(22 stones), against 20 stones for Lithidion and 17 for MyProgram. Finally,
Lithidion has the best hardware: a SUN HyperSparc 125 MHz, with 288 Mb RAM,
provided by the Vrije Universiteit, while MyProgram runs on a Pentium 133 MHz
with 48 Mb, and Marvin runs on PowerPC 601, 80 MHz and 16 Mb Ram.

Before the tournament, none of the competitors had a clue whose program would
be best, since they had not played any serious opponents for 3 years. Note that
except these three programs there seem to be no other Awari players (human or
machine) within their class: all human opponents the programs have had in
games against well-tested versions were defeated early in the game, with large
margins.

-------
DAY ONE
-------

So, on Saturday morning, 09.00 am, the tournament was started. The programs
meet each other 6 times, 3 times playing South and 3 times playing North, thus
12 games per program, and a total of 18 in the tournament. Each day, 6 of these
games are played.

At first glance, just looking at the results of the Saturday games,
the outcome of the tournament almost seems determined: MyProgram won all of
its games, while Marving beat Lithidion twice. So, the old champion was doing
badly.

If we look at the details of the games, things were less clear:

MyProgram vs. Marvin.

- In one of the games between Marving and MyProgram, Marvin sufferred from
  a time-control problem, leaving him to little time near the time control.
  The inferior moves computed with little time left, made him an easy target
  for MyProgram, at that stage searching on average TEN plies more than
  his opponent.
- In the other game MyProgram played the strong side (South), and shortly after
  Marvin left its book (move 12), MyProgram took the lead. At move 26 the win
  was secured.

MyProgram vs. Lithidion

- In the first game, MyProgram playing South, Lithidion was crushed. After 8
  moves in book, Lithidion lost on average 2 stones per move for the next
  5 moves. At move 16, Lithidion saw the loss. There was some confusion
  when the game was continued, since an initial 1 stone loss was enlarged by
  the seemingly mishandling of the position by a-b search, leading to 
  a predicted 12 stone loss. Then, when playing from the end-game database,
  the loss was decreasing all the time, back to a 2 stone difference.
  MyProgram even seemed to be unsure of the win after all. Finally all the
  humans saw what they had missed: MyProgram had already captured 25 stones
  and thus the game had been over at 22. Playing on after the end, lead to
  the confusing messages from the programs.
- The second game has been played after the two games between MyProgram and
  Marvin which gave Lithidion's authors a change to try to revive pn-search
  from the darkest part of the four-year old code. After quickly linking the
  seemingly untouched functions to the active code, and a short check of the
  vital signs of the program, they decided to give it a try. A-b search tried
  hard in the first part of the game, but could not help getting behind 6
  stones. Suddenly pn-search started to see past the horizons of a-b search 
  and had favorable messages, indicating that Lithidion could hope for a 
  draw or even more. While a-b search was unaware of this, MyPrograms +6
  lead dropped in his perspective to 0, confirming pn-searches claims.
  Unfortunately, only a few moves later pn-search, started talking rubbish:
  First it said that it would win the game, only 1 ply later, while checking
  the options for the opponent, saying the same for him. Needless to say, the
  buggy software helped securing a second victory for MyProgram.

Marvin vs. Lithidion

- In the first game, Marvins book played 19 moves for him, a tournament
  record. After these moves, Marvin had a one-stone lead. Somewhere in the
  next five moves, Lithidion must have erred, as quite suddenly, while being
  only 3 stones down, Lithidion found a loss, confirmed by Marvin.
- The second game against Marvin, the last of the day, left Lithidion's
  authors in a dilemma. We had found a collection of bugs, and did not know
  what to do. One thing was sure: the 20-stone database contained some (small?)
  errors. At that stage we even suspected the database being responsible for
  the pn-search behavior. Also, while they distrusted pn-search, it had
  given them hope in the second game against Lithidion. So, it was decided
  to play using an old, perfect, 17-stone database. During play, one of the
  authors was checking the code and found that the blame was placed
  incorrectly. Not the database was to blame, but pn-search. At the time in
  the game we needed pn-search most (to prevent a blunder found by a-b search),
  it failed, and a quiet position turned into a disaster, followed by a
  quick loss.

So, as all experienced game programmers will understand, our
saturday night fever did not lead to visits to Amsterdam's night live.
Instead, Eric went home (nothing much too improve), and Thomas, Maarten en
Victor stayed to work on their programs, until 04.30 am.

-------
DAY TWO
-------

After a well-deserved 4 hour sleep, Thomas, Maarten en Victor were ready for
another day of battle. First, Marvin and Lithidion had to battle it out, with
Lithidion playing the South. The opening went as expected: Lithidion getting
out of book earlier then Marving (move 4 against move 10). Things seemed even
at move 10. From then on, Lithidion slowly but surely started to take the
lead. Things got exciting at move 23, when pn-search warned that Marvin should
better not play move C, as it would lose, while up to 20 ply Marvin thought
that C was the best move. Marvin had time to finish 21 ply, and switched to
move B, leaving the game as yet undecided. However, at move 27, there was
no escape: pn-search saw a win. At move 30, Marvin saw the same, and after
a few minutes of stress in which both programs played almost instant moves,
we found that pn-search this time had not betrayed Lithidion. Lithidion had
scored its first point.

The second game proved to be the real test: playing against MyProgram, while
MyProgram played South. Things happened so fast, that the game was almost over
before we knew what had happened. In slow-motion it looked like this:
The first 11 ply were from book, with a small lead (between 1 and 2 stones)
for MyProgram. From move 7, where MyProgram searched 3 plies more
(26 against 23) than Lithidion, to move 11, where Lithidion had caught up
and searched 25 ply, Lithidion saw a zero-score, while MyProgram, using its
smaller database, still thought it was 2 stones ahead. In two more moves,
Lithidion started outrunning MyProgram in plies, not in stones: using a 51-ply
and 53-ply search, Lithidion's a-b search showed that the position was a draw,
which was confirmed by pn-search. MyProgram suspected the same, and at move
16 it was all over: the first draw of the tournament, and the first time
MyProgram had to accept anything but a win.

This is where we are at the moment, while the third game is under way, with
a small lead just after the opening for Marvin.

We keep you posted.

Victor Allis.
