Dragon 3.0
GM Larry Kaufman will introduce Dragon 3 by Komodo Chess.
We have released Dragon 3 at komodochess.com. Dragon uses NNUE (Neural Network Updated Efficiently) technology, originally developed for the game of shogi. Komodo has a great deal of chess knowledge in its evaluation. Training an NNUE network based on this evaluation was both an advantage and a challenge, requiring experimentation with architectures and data generation of billions of positions. The reinforcement learning phase for Dragon is in its infancy, but is already showing great promise.
Dragon 3 is a huge strength improvement over Dragon 2, roughly 80 elo at blitz on eight threads and 107 elo on one (using normal openings), about 132 elo in MCTS mode, and 148 elo when playing Fischerandom (chess960). It is a significant improvement over Dragon 2.6.1, about 17 elo on one thread blitz, about 20 elo in MCTS mode. The gains over Komodo 14.1, the last pre-dragon release (Nov 2020), are in the 300 to 400 Elo range on one thread, depending on mode and game type, all at CCRL blitz time control (2' + 1").
We have released Dragon 3 at komodochess.com. Dragon uses NNUE (Neural Network Updated Efficiently) technology, originally developed for the game of shogi. Komodo has a great deal of chess knowledge in its evaluation. Training an NNUE network based on this evaluation was both an advantage and a challenge, requiring experimentation with architectures and data generation of billions of positions. The reinforcement learning phase for Dragon is in its infancy, but is already showing great promise.
Dragon 3 is a huge strength improvement over Dragon 2, roughly 80 elo at blitz on eight threads and 107 elo on one (using normal openings), about 132 elo in MCTS mode, and 148 elo when playing Fischerandom (chess960). It is a significant improvement over Dragon 2.6.1, about 17 elo on one thread blitz, about 20 elo in MCTS mode. The gains over Komodo 14.1, the last pre-dragon release (Nov 2020), are in the 300 to 400 Elo range on one thread, depending on mode and game type, all at CCRL blitz time control (2' + 1").
The improvement over Dragon 2 is due to a larger, "smarter", and better-trained net, to deeper search due to search enhancements, and to parameter tuning. The improvement from Komodo to Dragon is due to the embedded neural network providing much more accurate evaluation and also in some sense gaining an extra ply or so of search by seeing some tactics that a normal eval can not see. Our testing also shows that Dragon is especially strong in Fischerandom (960) chess, which was confirmed when Dragon won the TCEC FRC 2021 championship.
Dragon also won the 2021 TCEC Swiss and Swiss2 championships (normal chess). We believe that Dragon will play in a more human-like style than standard Komodo since it relies on learning what actually works in games rather than just on pre-assigned values for eval terms.
Comments
Post a Comment