kobanium / aobazero

Aoba Zero
Other
54 stars 8 forks source link

AobaZero (AlphaZero Shogi test) has finished. #53

Open yssaya opened 3 years ago

yssaya commented 3 years ago

Since March 2019, we have generated 39.8 million games for 2 years and 2 months (AlphaZero has generated 24 million games). We would like to thank everyone who has cooperated in generating games, and those who have reported bugs and played with.

AobaZero is estimated to be 154 Elo weaker than AlphaZero (377 Elo in Sente(Black), 86 Elo in Gote(White), see below).

From only Shogi rule, it has rediscovered most of the Castles and Openings that Shogi players often use.

Confirmed Castles  : Yagura, Gangi, Mino, High Mino, Silver Crown, Left Mino,
                     Central House, Right King, Yagura Anaguma
Confirmed Openings : Double Wing Attack, Side Pawn Capture, Side Pawn Capture Aono Style,
                     Bishop Exchange Climbing Silver, Bishop Exchange Rushing Silver,
                     Bishop Exchange Reclining Silver, Gote 4th File Rook,
                     Yagura Waki System, Twisting Rook

In Sente, it does not play Ranging Rook. But in Gote, it preferred to play 4th File Rook. It does not play orthodox Anaguma(78 Gold, 79 Gold), but plays Yagura Anaguma(78 Gold, 67 Gold). It haven't found a variety of Ranging Rook(Central Rook, 3rd File Rook, etc).

The latest playing style is the immediately hitting static Rook party, which prefers Double Wing Attack and Bishop Exchange, which are similar to AlphaZero's published 100 games.

The advantages are an evaluation of the delicate arrangement of pieces in the early stages, and an evaluation of the possibility of Entering King. Especially with other software opponents, there are many declaration wins by Entering King. For example, against elmo(WCSC27), 97% of Sente wins are declaration wins. The weaknesses are missing some tactical moves in end-game, and overlooking long distance Rook and Bishop control.

AobaZero winrate against elmo(WCSC27), 330 games, AlphaZero 10 seconds/move (Equivalent)
Winrate       0.688 (219 wins 16 draws 95 losses)                        diff 137 Elo
Sente winrate 0.748 (120 wins  7 draws 38 losses) 116 declare wins(97%)  diff 189 Elo
Gote  winrate 0.627 ( 99 wins  9 draws 57 losses)  56 declare wins(57%)  diff  90 Elo

*AobaZero 580k playouts/move, elmo 251000k nodes/move (6 threads)

AlphaZero winrate against elmo(WCSC27), 1000 games, 3 hours per each.
Winrate       0.918                                                      diff 420 Elo
Sente winrate 0.982                                                      diff 695 Elo
Gote  winrate 0.853                                                      diff 305 Elo

*AlphaZero uses four 1st generatrion TPUs(58k playouts/second), elmo uses on Xeon 2.2GHz 44CPU, 44 threads(25100k/second).

AobaZero's weight is w1650.txt, it is about 40 Elo weaker than latest. elmo(44 threads) is 89 Elo weaker than elmo(6 threads), when it searches fixed 25100k nodes. (Comparing against Kristallweizen(16 threads) fixed 5500 nodes per move.)

Average  154 Elo (420 - 137 - 89 - 40)
Sente    377 Elo (695 - 189 - 89 - 40)
gote      86 Elo (305 -  90 - 89 - 40)

Elo graph, game records and weights. http://www.yss-aya.com/aobazero/index_e.html Sample games without noise. http://www.yss-aya.com/aobazero/no_noise/sample.html