Chrysaora may very well be was the first serious attempt to build a world class chess engine in Nim.
Chrysaora started as an attempt to use supervised learning and a small image classification styled network to play chess using Python. The results of this experiment are stored in chrysaora_py. The project has now shifted to building an experimental "hybrid" chess engine. This hybrid will have an evaluation function that consists of two parts:
- A reinforcement learning trained neural network
- A handcrafted element, used for quiescence searching and depth-1 move ordering.
Or at least, that's the plan...
The work on chrysaora is heavily inspired by and based on the Giraffe and KnightCap papers, both of which I consulted liberally while writing the engine. Most of my training code only came together after really understanding Knightcap.
Friend, Chrysaora can barely even play competent chess, let alone good chess.
Learning and evaluation is done using arraymancer, which is the only dependency outside the Nim standard library. You must install arraymancer at version 0.6.1 and no later. Why? Bugs mostly.
Chrysaora is named after the genus of jellyfish, which in turn is named after Chrysaor, a being from Greek mythology. Chrysaor roughly translates as "he who has a golden armament." Major releases of Chrysaora are codenamed after species (as opposed to genuses) of jellyfish, typically things I find cool.
For some reason I use female pronouns when referring to Chrysaora in my head, but I can't imagine she particularly cares considering she's a nonsentient chess engine.
- v0.1.x Noctiluca - Named after Pelagia noctlicua, a bioluminescent jellyfish. There was a bioluminescent algae bloom while I was coding Chrysaora. Everyone was quarantined and I coded Chrysaora instead of going to see it. Sad.
Each weights file is labeled by its major version, a training iteration delimiter, and the number of games that went into that training run.
For example:
noctiluca-t1-20.txt
This weights file uses the network defined by the base version noctiluca and run through one training run (t1). The 20 denotes that this training run used 20 games of play to update the weights.
t0 is specifically reserved for bootstrapped weights.
Chrysaora will be supported until I get my PhD (in Physics), it wins a season of the TCEC, or Nim dies, whichever comes first.
The easiest way to build Chrysaora is to use nimble, which will handle the requirements and dependencies for you:
nimble build -d:danger
You can also self build. Chrysaora requires a minimum Nim version of 1.2.0. With Nim (and [email protected]) installed, you can instead build Chrysaora with
nim c -d:danger src/chrysaora.nim
However, if you want to build Chrysaora and then train it yourself you must have the Boehm GC installed and instead build as
nim c -d:danger --gc:boehm src/chrysaora.nim
This is because Chrysaora will crash the default GC when on Nim >= 1.4.6. Why? I don't actually know. In all honesty, Chrysaora pushes Nim pretty hard, both in and of itself but also through Arraymancer, which also tends to push Nim to its limits.
In order to train Chrysaora, you must build a version that can support training. Currently all versions of Chrysaora build with training built in. However, if you try and build it with the default GC you will segfault the GC. Sorry. See the Building section for details on how to work around this.
Chrysaora learns by playing chess games against an opponent. It is possible to self train Chrysaora by pitting two engine instances against each other, but the weight updates of one engine will overwrite the other. I would recommend providing a file of openings to start from, as Chrysaora will play subsequent games very similarly if she is allowed to pick the opening moves every time.
Currently Chrysaora will train by setting the in game uci option to true. This is done by running the following after loading an instance of Chrysaora but before telling her to find a move:
setoption name Train value true
If you use cutechess-cli to run Chrysaora against another engine to train against, you must add option.Train=true to Chrysaora's command in order for training mode to be turned on.
While training, Chrysaora will keep a running record of its internal evaluations as well as the gradients used to calculate these evaluations for each of its own color board states. I.e. if Chrysaora is playing white, then it will store evaluations and gradients for all white moves in the game (as these are the moves that Chrysaora herself makes.)
At the end of the game Chrysaora will use the TDleaf(lambda) update rules to update the internal weights based on the game performance. This is done when either ucinewgame or quit is passed to Chrysaora.
Chrysaora will save the weights after every 10 games and upon exit. You can change the number of games between saves by modifying the save_after variable in train.nim.
Chrysaora can bootstrap her weights to approximate its own internal handcrafted evaluation function. This can be done by running bootstrap x from the uci interface at any time once the engine is initialized, where x is the number of training epochs to use for bootstrapping. The bootstrap weights will be saved as majorversion-t0.txt. Chrysaora can train from random weights, but by bootstrapping and putting herself near a good minimum in parameter space the training time to get "good" is massively reduced.
Chrysaora bootstraps by pulling two game states from pgns that are saved in /games/train/. These game states are passed to the network both in their original form and a colorswapped version. This helps Chrysaora learn to play both colors on the chess board.
- Move Generation
- Fancy Magic Bitboards
- Redundant Mailbox representation
- Hand Crafted Evaluation
- Centipawn piece imbalance
- Piece-square tables
- Network Evaluation: 111 input features
- Number of each piece (10 features)
- Side to move (1 feature)
- Castling rights (4 features)
- Square position (64 features)
- Piece existence (32 features)
- Search
- Fail-hard alpha-beta pruning minimax
- Zobrist hashing indexed transposition table
- Iterative deepening
- Handcrafted move ordering for depth <= 2 plies
- Ability to build handcrafted evaluation versions of Chrysaora
- Self training
I'd like to say a very special thank you to the following engines, which I consulted during the coding of Chrysaora:
- Ethereal, for help with understanding bitboards, as well as their magic numbers
- Laser, for the idea of recording the ep target square between moves
- Stockfish, for help understanding magic number generation
- Lc0, for paving the way for NN engines
- Giraffe, for providing the initial network structures for Chrysaora
Additionally I'd like to thank the Chess Programming Wiki for its help in getting the project started.
Neural Network based chess engines seem to be the "new era" in computer chess, especially so since Lc0 just won the superfinal in S15 of the TCEC. However, NN based engines come with their own set of problems. The general width and size of the NN required to play chess at the level that LC0 plays at is massive, explaining a major part of why NN based engines tend to be so much slower than traditional hand crafted engines. Consider that if Lc0 could achieve the nps that Stockfish does it would be, in my opinion, unbeatable.
By hybridizing a NN with a handcrafted evaluation function, I hope to achieve similar playing performance to top level engines with a much narrower network structure. The much narrower network should compute faster, and if my coding and optimization prowess is up to the task, achieve a potentially higher nodes per second.
On the other hand it could blow up in my face. So far it's just blown up in my face.
My network-handcrafted evaluation hybridization consists mostly in using a handcrafted eval where the robustness of a full neural network isn't required. In nearly every case except for extremely contrived examples a handcrafted eval will be faster than a network.
Chrysaora uses a handcrafted evaluation to order the moves at depth 1 in the iterative deepening framework. Quiesence searching is also done with the handcrafted evaluation. Quiesence searching is essential to ensure our search tree doesn't end in a "noisy" move. By using a handcrafted evaluation for this, I hope to speed the high depth probes of Chrysaora to be fast enough that in the same amount of time a neural network quiesence search is completed I can search to an additional depth with a handcrafted quiesence.
Additionally, unlike most (if not all?) highly competitive NN engines, Chrysaora uses an alpha-beta pruned minimax search rather than a Monte Carlo Tree Search.
You can find the details of Chrysaora's license in the LICENSE.txt file. Chrysaora is licensed under GPL-3.