IEEE Big Data Cup 2024

Help us create better chess puzzles!

PuzzleChess botSoftware Development
Volunteers needed for a chess AI project

We are seeking volunteers to solve our puzzles and help promote AI research to create chess study materials.
Go to and follow the instructions.
Your help is much appreciated!
Continue reading to learn more about this initiative!

You might also be interested in our puzzle prediction competition with financial prizes if you are a data specialist! See for details.

Solving puzzles is a great way to repeatedly expose yourself to various scenarios that could happen in your chess games. If you have a particular weakness in your play, e.g. you struggle to create mating nets or you miss pawn breaks, puzzles are a way to work on it in a controlled environment.

But where do the puzzles come from? Well-known trainers such as Aagaard and Dvoretsky create targeted puzzle sets by analysing their games, browsing databases, reading books and articles, and sometimes composing problems from scratch.

A sample card from Mark Dvoretsky’s famous collection of card index files. Nowadays, such collections are typically stored digitally
Source: chessbase

Through many years of curating their growing collection, having their students work through them, discussing and playing them out, coaches create a true chess training treasure. They learn which puzzles are beginner-friendly and which are only suitable for very strong players. They understand where and how different topics intertwine and are able to accurately target their pupils’ weaknesses.

Things are more difficult if you don't have access to a coach.

Lichess generates puzzles using algorithms (see this blog post for details) and estimates their difficulty using the Glicko rating. In short, whenever you try to solve a puzzle, lichess treats it as a game. If you solve the puzzle, your puzzle rating increases and the puzzle’s difficulty rating decreases. Puzzle’s rating usually stabilises after around 20-30 attempts.

Would it be possible to estimate puzzle difficulty just by looking at the position on the board and the solution, without relying on crowd-sourced attempts at solving it? Intuitively, some features can help us make a prediction. For example, a mate in three is usually more difficult to find than a mate in one. Could various features like that be identified and codified into an algorithm? Perhaps we could employ Artificial Intelligence to help with that?

We are a group of AI scientists from Poland who would like to see more research done in this area. With that in mind, we launched a data competition as a part of the IEEE Big Data 2024 conference to predict chess puzzle difficulty, with cash and conference publication prizes to attract the best AI talent.

However, we still need help from chess aficionados such as yourself! Data competitions require a substantial set of previously unseen examples to ensure a fair verdict. We’ve deployed a clone of the lichess server where the puzzles for our competition are available. Now we need as many people as possible to go there and try their best at solving the puzzles for them to serve as a test set The more you solve, the better!

But wait, there is more! Additionally, if you are a data-savvy person interested in participating in the competition, you can find it here: Good luck!

Thank you for helping us make chess training more accessible to everyone!