The Best Leela Chess Zero (LC0) Neural Nets for Chess Enthusiasts

As of August 2023, Leela Chess Zero (LC0) has several neural network versions that are notable for their performance. These neural nets, which have evolved through various versions, continue to enhance the capabilities of LC0 in chess competitions and general usage. This article explores some of the best-known neural nets used in LC0, offering insights into their strengths and recommended usage.

Overview of Key Versions

Several notable neural net versions of LC0 have been developed and refined over time. These versions incorporate improvements in evaluation functions, training methodologies, and hardware optimizations. Some of the most prominent versions include:

LC0 0.30

This version was particularly strong in various chess competitions due to several improvements over previous versions. It introduced enhancements in its evaluation function and training methodology, making it a widely recognized choice for chess enthusiasts and developers alike.

LC0 0.32

The 0.32 version brought more sophisticated training techniques and larger datasets, leading to improved playing strength. Optimizations in this version allowed for more effective use of hardware, further enhancing its capabilities.

LC0 0.34 and Later

Subsequent versions of LC0 continued to build on the foundation of earlier models, incorporating lessons learned from match play and user feedback. These improvements included updates to the training process and architecture enhancements, making LC0 a robust choice for serious chess players.

Custom Networks

Many users and developers create custom networks based on the LC0 framework, tweaking parameters and training on specific datasets. These custom networks can outperform standard versions in particular scenarios, offering unique strengths that cater to specific needs.

Community Contributions

The LC0 community frequently shares their trained networks, which can vary in strength depending on their training conditions and the hardware used. Some of these community models have gained recognition for their unique strengths and have become valuable additions to the LC0 ecosystem.

Evaluation Criteria

When evaluating neural nets, several factors come into play:

Performance: Assessing the neural net's playing strength in competitive settings. Size: Different neural net sizes are recommended based on their intended use. Smaller nets (less than 10b) are suitable for sparring against humans, while larger nets (24b or more) are better suited for more intensive tasks. Hardware Compatibility: Some neural net versions are optimized for specific hardware, such as RTX GPUs, providing better performance. User Feedback and Match Play: Feedback from users and results from match play are crucial in determining the effectiveness and reliability of a neural net.

Network List

Here are some of the most prominent networks:

SV-30b-t40-1573

Trained on T40 data, this network is part of the Sergio-V repository. It has shown strong performance in various chess competitions and is recommended for running on RTX GPUs.

Latest T60 run 1 networks (24b x 320f)

These networks are currently part of the main run and are trained on T60 data, making them a suitable choice for a wide range of chess activities.

J13B.2-136

Trained on J13B.2 data, this network is also available on GitHub and is recommended for running on non-RTX cards or for tasks where the time control (TC) is on the order of seconds.

Strength Comparison

Some specific comparisons between neural nets have been made. For instance, a match between a 384x30b model and a 256x20b model shows:

Sergio_384x30_1573 vs Sergio_256x20_1467: The score 24 - 59 - 117 indicates a significant advantage for the 384x30b model, with an Elo difference of -61.4. Match results: 611 - 541 - 848 [0.517], indicating a clear advantage for the 384x30b model.

Conclusion

Choosing the best Leela Chess Zero neural net depends on your specific needs and the anticipated hardware. Whether you are looking for a general-purpose neural net or a highly specialized one, there is a version of LC0 to suit your requirements. The LC0 community continues to produce and share models, enhancing the overall performance and capabilities of the platform.