Black Marlin
WIP UCI Chess Engine
![]() |
---|
Logo by Alex Brunetti |
Black Marlin is a UCI compliant chess engine fully written in Rust by Doruk Sekercioglu.
Black Marlin supports Chess960 as of December 26, 2021 although no extensive testing has been done.
Make sure to have Git LFS and compile the engine with make
in the root directory. This will output a BlackMarlin executable.
Black Marlin doesn't come with a built-in GUI. The recommended way of playing against the engine is to get the latest release or compile it locally and use it along with a Chess GUI that supports the UCI protocol.
The repository used for NN training is NNUE Marlin.
The above repo has been replaced by Marlinflow.
Black Marlin Search
- Iterative Deepening
- Aspiration Windows
- Principal Variation Search
- Lazy SMP
Black Marlin Evaluation
Black Marlin Time Management
- Search Evaluation Instability
- Failed Searches at Root
- Best Move consistency
Iterative Deepening
Iterative Deepening is the idea of performing consecutive searches at higher depths each time.
Aspiration Windows
Aspiration Windows are a way of reducing the search space by making the assumption that the resulting evaluation will be in [alpha, beta]
range. If we get a score such that alpha < score < beta
, we know that the resulting score from the search is exact and thus can be used. Otherwise, we execute another search with a larger window.
Principal Variation Search
In Principal Variation Search (PVS), we do a search with the window [alpha - 1, alpha]
for each move to see if it is worse than alpha efficiently. If so, we don't need to execute a full search on the given move.
Pruning
Pruning in general is the idea of ignoring certain parts of the search tree in order to reduce the amount of time it takes to execute a search.
Reverse Futility Pruning
Reverse Futility Pruning (RFP) is the idea of returning the evaluation of the current position if it is above beta by a margin.
Null Move Pruning
Null Move Pruning (NMP) is based on the assumption that, if we can reach beta by not making a move at lower depths, we will most likely be able to reach it anyway.
Late Move Pruning
Late Move Pruning (LMP) is the idea that all quiet moves can be pruned after searching the first few given by the move ordering algorithm.
Futility Pruning
Futility Pruning is the idea that quiet moves don't tend to improve positions significantly, therefore we can safely prune quiet moves in positions where the current position is evaluated to be below alpha by a margin.
History Heuristic
History Heuristic is usually a table that keeps track of which moves failed high/low in the past. This can be used to identify moves that tend to be good in the current search space.
History Based Pruning
History Based Pruning is the idea of using the history table to see which moves performed worse in the local search area and pruning accordingly.
SEE Assisted Futility Pruning
Doing Futility Pruning at higher depths by correcting the evaluation based on static exchange evaluation (SEE).
Late Move Reductions
Late Move Reductions (LMR) are a way of proving if a move is lower than alpha quickly. This is done by searching moves that are expected to underperform at a lower depth.
History Based Reductions
History Based Reductions are the idea of adjusting LMR reductions based on the history score of a move. This is done for both quiet and noisy (non-quiet) moves as history is kept separately for both.
Other Reductions
Black Marlin additionally reduces less in principal variation nodes and nodes that improve the static evaluation of a position compared to its parent.
Extensions
Usually the search algortihm is optimized by pruning or reducing where moves aren't promising. However, it can also be done by executing a deeper search on promising moves.
Check Extensions
We don't want to miss three-fold repetition lines or forcing sequences, so we extend moves that give a check.
Singular Extensions
If a move is better than all other moves by a margin, we can extend the search on this move. Candidate singular moves are usually decided based on the transposition table.
Multi Cut
If a move isn't singular and the singular move is above beta, we know that there is more than one move that beats beta. We can thus fail soft and return the singular move beta.
Low Depth Singular Extensions
This idea is inspired by Koivisto however the implementation is slightly different. At lower depths, Black Marlin doesn't search to prove singularity but rather uses the static evaluation directly. Since static evaluation isn't a reliable estimate, Black Marlin doesn't apply multi cut pruning if low depth singular extensions are done.
Move Ordering
In order to maximize the efficiency of alpha beta search, we optimally want to try the best moves first.
Move from Transposition Table
The Transposition Table (TT) gives us the move we have found to be the best in case we have visited the position previously. TT moves are one of the most crucial ways of improving move ordering.
Winning Captures
We can evaluate each move by using SEE and put winning captures first in move ordering. Although moves aren't ordered by SEE but rather by their capture history scores and the value of the piece the move is capturing. This both provides more performance over pure SEE and benefits from the adaptive nature of history.
Killer Heuristic
Killer Heuristic is the idea of using quiet moves that produced a beta cutoff in sibling nodes. This is based on the assumption that a move doesn't change the position a lot and we can benefit from moves that worked in very similar positions.
Refutation Move
Black Marlin keeps a table of the last best quiet response to each move. These moves are later used in quiet move ordering.
History Based Quiet Ordering
We can order quiet moves based on their history scores in order to search quiets that historically performed better than others first.
Bad Captures
Bad Captures are moves that most likely make the position worse. They are usually put last as they rarely improve the position. Similar to winning captures, bad captures are also sorted by capture history.
Under Promotions
There isn't really a reason to promote to a piece other than the queen in most positions. All promotions other than queen promotions are put to the very last of the move ordering list.
Lazy SMP
Lazy SMP stands for Lazy Symmetric Multi Processing. The idea behind Lazy SMP is to execute the search function on N many threads and share the hash table. This allows for faster filling of the hash table, resulting in a faster and wider search.
Atomic Keys and Entries
In order to prevent data races and data corruption, we can use the "XOR Trick". Instead of saving the hash of the position directlly, we XOR it with the related entry. When we access the table later on, we can revert the hash back with the corresponding entry. This ensures a different hash will be produced in case the entry was corrupted.
Hand Crafted Evaluation
Prior to NNUE, most engines used a hand crafted evaluation function. Hand crafted evaluation (HCE) included hand picked features such as material, pawn structure, piece squares and similar. These values quickly get overwhelming to tune, in order to solve this problem, most engines employ a strategy called "Texel Tuning". The idea behind Texel Tuning is to optimize the evaluation function in such a way that it'll reflect the result of the game the best. Once we obtain a loss function, we can use metaheuristics such as genetic algorithms or gradient based algorithms such as AdaGrad to tune the hand crafted evaluation function.
Neural Networks
A Machine Learned model that has information on all the board is naturally expected to perform better than hand picked features.
Training the Neural Network
There are many distinct ways of training a neural network for chess however the one that is employed the most is using the evaluation score from a low depth search. This results in evaluations that see more into the future and produce neural networks more capable than HCE functions.
Data Generation
Data is usually generated by self-play where the engine plays against itself at a fixed depth/node count. This enables us to also keep track of win draw loss (WDL) information to be later used in training.
Training
In order keep the weights low and make the neural network more aware of evaluations close to 0, the evaluation score is usually scaled and passed through sigmoid. Usually using a formula along the lines of sigmoid(eval * SCALE)
. We can also add a WDL term to our train set in order to make the networ more aware of how winnable positions are. This can be done by taking the weighted average of WDL and evaluation.
Efficiently Updatable Neural Networks
Neural Networks are the state of the art machine learning models. Although they have one problem: they are usually very slow in execution. Efficietly Updatable Neural Networks (NNUE) solves this problem by doing incremental updates.
Incremental Updates
Alpha-Beta Search has one feature that allows us to efficiently update our neural networks. Since a depth first search is employed, a series of moves are played one after the other only causing slight changes in the actual input. It is possible to take advantage of this and only calculate what has changed in the network than propogate forward completely.
Skipped Connections
Skipped connections are the idea of connecting a layer to a layer other than the next. This increases learning speed and the strength of the resulting model.
Buckets
A Neural Network might have sub networks that it picks from based on a condition. This enables the network to better distinguish between different types of positions, usually decided by game phase(mid game, end game). Buckets cause almost no slow down in the resulting neural network while increasing the representative power of the model.