Lc0-stockfish-hybrid is a (mostly) UCI-compliant chess engine based off two separate projects, Stockfish the leading traditional chess engine, and LC0, or leela the leading open source neural network chess engine based off AlphaZero.
There are detailed build instructions in INSTALL.md. If you need windows or linux binaries, you can contact me at [email protected].
Both the Lc0 chess engine and the stockfish engine are extremely strong, and have been dramatically improving due to improved heuristics for stockfish, and improved training for Lc0. However, they are very different engines, based off very different methods and strengths, and so the goal is to combine those strengths.
- In head to head matches, typically Lc0 wins reliably, but Stockfish is better at beating weaker engines (avoiding draws more successfully), so Stockfish still wins many engine tournaments.
- On the hardware in these tournaments, Stockfish evaluates around 60,000,000 nps, whereas Lc0 evaluates around 60,000. This is 3 orders of magnitude difference, yet they evaluate to a similar depth in most positions (around 35), indicating that Stockfish sees a much wider tree, better evaluating more tactics positions, whereas Lc0 sees a deeper tree, evaluating deep endgames and positional weaknesses using its vastly superior heuristic.
- Lc0 is not trained well for winning endgames. It can draw completely winning positions because it thinks every move is winning, including some that are not. In tournaments, a tablebase has to be used to mitigate this problem.
- Stockfish is relatively bad at openings. Opening books are not allowed in tournaments, and Stockfish's heuristic is not able to evaluate the complex middle game positions that openings produce.
- Lc0 is relatively bad at tactics, dropping pieces in short time controls.
So how to combine Stockfish's ability to win games and handle tactics with Lc0's ability to evaluate positions?
The core position evaluation algorithm is Lc0's Monte Carlo Tree Search (MCTS) Algorithm cloned from the AlphaZero paper. There is one change:
In MCTS selection, do not select moves that Stockfish has evaluated as bad.
This algorithm can be implemented much more efficiently than you would think. Contrary to appearances, Stockfish's core loop outputs a binary output. It does alpha-beta search where alpha and beta are the same number, so the only output of the core loop is whether the estimation evaluation is above or below alpha. To then compute the actual estimated value it reports, it does a binary search over candidate values. This central loop is extremely highly optimized, so using it directly is going to be extraordinarily efficient. This efficiency issue is a large part of the inspiration for the above algorithmic change.
In order to implement the algorithm completely, there are several aspects to consider:
- What is a bad move?
- How deep does Stockfish search to determine if a move is good or bad?
- What does Lc0 do if Stockfish has not determined whether the move is good or bad yet?
There is also a relevant design question: how does a seemingly crude binary criteria to explore edges in the tree produce high quality evaluations?
Moves that Stockfish thinks are bad are easy to characterize: A move is bad if it is 60 CP (centi pawns, a measure of material advantage) worse than Stockfish's evaluation of the root board position (this constant is a tunable hyperparameter).
To implement this efficiently, there is one thread computing the value of the root position, and many other threads evaluating whether the moves all over the tree are "bad" under the above criteria (or both players).
Stockfish's evaluation loop uses an iterative algorithm which slowly increases the search depth of the engine. Inspired by this approach, the algorithm does the following:
- All moves on all nodes that Lc0 explored at least once are evaluated to depth 3 (this is a tunable hyperparameter).
- Evaluation upgrades are given priority based off the following formula:
priority = Node Weight / Computation Time
This formula is designed so that fast evaluations are prioritized and evaluations which Lc0 weights highly (meaning it has searched the node many times) are also prioritized. Computation time is estimated to be the computation time of previos searches of the move.
This is simple, Lc0 assumes all moves are valid if Stockfish has not computed anything yet.
Why does this algorithm produce stable, high quality evaluations despite a seemingly crude exploration criteria?
One interesting feature of this algorithm is its builtin stability. If Stockfish rates a move as "bad" incorrectly, because of a shallow search depth, that does not prevent the node from being searched by Lc0, only the move. So if the node continues to be searched, the move which is weighted based off how often Lc0 explored the node, not the move, will continue to move up the priority queue until it is reevaluated by Stockfish at a greater depth.
Many similar variations on the algorithm I tried failed miserably because it did not have this self-correcting feature.
While it is hard to visualize a huge game tree, small parts of the game tree can be visualized to understand what the method actually does.
The numbers at each nodes in the graph (and the thickness of the borders) represent how many times the node is visited in the MCTS simulation (the number of subnodes that are evaluated). The red lines are edges which Stockfish has decided are definitely bad, and which will no longer be explored (but their values are still included in the evaluations of parent nodes, for evaluation stability). Nodes with less than 5 children are pruned.
Here is a evaluation graph in a sharp middlegame position where every single move is forced, the kind where Stockfish shows its best strengths.
Vanilla Lc0 evaluating 1000 nodes
Hybrid approach with Lc0 evaluating 1000 nodes
As you can see, adding the Stockfish signal completely changes the resulting evaluation tree. In particular, note that the tree searches much deeper, presumably because Stockfish's signal eliminates enough moves that it is more confident about where it should search.
Of course, Stockfish generally does make the best move in these sort of positions, so you might wonder why you can't just have Stockfish make the move, but the idea is that sometimes there might be two possible moves with very sharp subtrees that terminate in very positional endgames, and you still want Lc0 to evaluate the conclusion of different forced moved sequences.
Here is an evaluation graph of a very slow middlegame where stockfish thinks that at least 10 moves at each node for 5 move are OK, i.e., it has no strong opinions about the position. Lc0 seems to evaluate the position better, as the hybrid engine eventually beat Stockfish from this position, despite the signal from Stockfish being quite useless.
Vanilla Lc0 evaluating 1000 nodes
Hybrid approach with Lc0 evaluating 1000 nodes
In this case, the evaluation tree still changes, but much less than in the tactical position. Many parts of the subtree are similar, the depth of various parts of the tree are similar. This means that when Stockfish is not as strong in the position, it does not interfere with Lc0's normal evaluations very much.
I accidentally deleted all this data :( Working to recreate.
I accidentally deleted all this data too :( Working to recreate.
Out of maybe 100 games vs Lc0, in time controls from 3+2min to 15+10, I have seen it win maybe once. The draw rate is incredibly high, but in general, the hybrid method loses head to Lc0 very consistently. I suspect that is because the Lc0 evaluation of both engines guides it into positions it is comfortable in, meaning less tactically sharp ones, so the stockfish evaluation does not add much, and sometimes it gets in the way to some small degree.
Leela Chess is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
Leela Chess is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with Leela Chess. If not, see http://www.gnu.org/licenses/.
The source files of Lc0 with the exception of the BLAS and OpenCL
backends (all files in the blas and opencl sub-directories) have
the following additional permission, as allowed under GNU GPL version 3
section 7:
If you modify this Program, or any covered work, by linking or combining it with NVIDIA Corporation's libraries from the NVIDIA CUDA Toolkit and the NVIDIA CUDA Deep Neural Network library (or a modified version of those libraries), containing parts covered by the terms of the respective license agreement, the licensors of this Program grant you additional permission to convey the resulting work.