There has been a lot of heated discussion surrounding the recent Google chess AlphaZero paper. Google published a manuscript where they describe their program learning chess from the basic rules, and improving through self play and a general learning framework. They claim to have reached a level of play that can defeat Stockfish 8 (not the most recent version) after a few hours of training. In my opinion this achievement is incredible, regardless of the criticism expressed in the chat (not a fair competition because the hardware Google used for each program was very different, Stockfish was not configured optimally, a few hours on Google hardware is equivalent to years for Stockfish testers). I for one would love to see a similar contestant in TCEC.
In game 81 there were no pawn exchanges after the opening. Houdini pushed pawns on the king side and evals became slightly negative. Komodo's queen went on a risky journey, capturing a pawn and creating a passer on the queen side but giving a rook for a bishop. Houdini had to control the white passer, this allowed Komodo to exchange down to a drawn queen ending. There were many exchanges in game 82, evals got as high as 0.5 but then dropped to 0. The game reached an opposite color bishop ending, and after 35 moves the engines agreed to a draw.
Game 83 started with a small eval advantage for Komodo. All knights were exchanged early in the game. Komodo was a pawn up when it chose to give its queen for a rook and bishop.
Evals stayed low while Komodo pushed its e pawn to create a strong bishop and pawn pair on the 6-7 ranks. Komodo limited the possible squares available for Houdini's queen until it was almost trapped in the corner. Komodo's eval rose above 1, Houdini was still calm.
The board was full of traps, Komodo's pieces coordinated perfectly. Komodo's eval was close to 3 before Houdini realized it was in deep trouble. The black queen was out of the corner and even captured a pawn. In order to let the queen get back and help the defense Houdini weakened the king side pawns.
Now Houdini's eval jumped, Komodo closed in on the black king and Houdini allowed the white passer to advance. Houdini took the passer and exchanged a pair of rooks but was about to lose a piece when the game was adjudicated.
In game 84 there were many exchanges. After queens were exchanged evals dropped to 0 and stayed there for the rest of the game, reaching a drawn rook ending.
Komodo had a small eval advantage in game 85 even though its king stayed uncastled in the center. All knights were off the board and with many pawns the remaining pieces had limited mobility. After queens were exchanged only RB vs RB remained and evals dropped to 0. Houdini gave two pawns to be able to place its rook on the 2nd rank, the game was adjudicated and the PV showed a series of exchanges and a drawn rook ending. In game 86 the center was open and both engines castled. Komodo gave a rook for a knight and got a pawn majority on the queen side. Evals dropped as the engines exchanged pieces. In a QR vs QB position and an open position the queens posed many threats, finally Komodo forced a draw by perpetual check.
Komodo started with an eval advantage in game 87, and evals gradualy increased. The center and king side were mostly blocked, both kings were apparently safe, Komodo had more space and a bishop pair that could explain the eval.
The engines shuffled for a while, evals continuing to increase. Komodo then pushed the c pawn, disrupting Houdini's pawn structure and threatening the weak pawn on e6. As a result Komodo was able to create a hole in the center and a passer, with evals over 2.5.
The e passer was well supported and threatened a black rook. Houdini tried to use a pin so Komodo moved the queen out of the way. After a series of exchanges most pieces including queens were removed, Houdini gave a piece to get rid of the white passer.
The rest of the game was straightforward, Komodo exchanged rooks and the piece advantage was enough for a win.
In game 88 the center and king side were again mostly blocked. This time black castled on the king side and the queen side pawns for both engines had more freedom to move. The engines exchanged most pieces through the open queen side until only RN vs RN were left.
Komodo had a queen side passer and material was equal, yet Houdini had an eval advantage of 0.7 and gradually increasing. Komodo had no counterplay and had to defend its weak pawns, in particular the backward e pawn. Houdini's knight replaced the rook in blocking the a passer, and this allowed the rook to get behind the black pawn line.
Komodo's king and knight could not move for fear of losing the e pawn. Houdini's king started the long walk around to the queen side. Once it was covering the a pawn the knight joined the attack and it was game over for Komodo. The black pawn line crumbled and a few moves later the game reached a won rook ending.
Komodo started a pawn up with a queen side passer in game 89 (Benko gambit). A long series of exchanges left only BB vs BN with equal pawns on move 26. Both engines' evals were close to 0, the bishop pair was not sufficient to push the white passer forward, game ended in a quick draw. Houdini kept the pieces on the board in game 90, with evals increasing to about 0.8. Komodo used its major pieces to defend the queen side, and with the rest of the board static the engines shuffled for a while with evals close to 0. The engines pushed a few pawns forward, opening files and exchanging pieces, as a result the queen side threat was removed and the game reached a tablebase draw.
No comments:
Post a Comment