personal experience regarding "self learning"

[deeds]

...What learning method works best for training with opening books ?...

Frankly, there is no difficulty in training with opening books, it's like training against Stockfish or any non-learning engine.

Books all show a static playing style + they have less variety because at best their authors only configure 2 or 3 moves/position + they handle a lot less openings than experience files Blessed bread for learning engines at training with them

For training with opening books, I use the same method as with non-learning engines.

Until now it works well, here are some results :
- with opening books
- with Stockfish 14.1
- with Stockfish 15.0

Recently I even tested it against a learning engine (set by default in tourney mode) but which had not been trained on the opening's list used by the daily tester it was like training against Stockfish or any non-learning engine. It worked well here too.

[deeds]

...How many rounds are needed for training with opening books ?

Except for round 0 (=100 games from the starting position) which is just used to discover what will be the playing style of the opening book, it often takes 2-3 rounds.

The opening books all do the same : they seek to take the opponent into little-known positions so that he falls into a trap or loses by time forfeit. Of course, its author took great care to protect his opening book against his own pitfalls.

So it works very well against an opponent who also has a static playing style. He will lose almost every round and will not learn the wards used by the opening book.

On the other hand, against a learning engine, after 1 or 2 rounds, it will have memorized the trapped lines and will avoid them by trying other moves (=training mode), often by making the book come out of its openings (=out-of-book). It becomes fun if the learning engine had already been trained on some of the openings corresponding to the playing style of the opening book.

[deeds]

Hey DeeDs, a little gift for you !

From your LS1's results :

Ah, it's sure that, each day, engine's trainers see bigger ELO gains than on fishtest but it's not really comparable.
Not the same vision of elo gain either among daily testers.

Thanks !

[deeds]

In conclusion, creating an experience file for a UCI chess engine is a multi-step process that integrates data collection, meticulous file management, and advanced learning techniques to produce a self-improving chess AI. The methodology outlined in this article based on the comprehensive guide available at deeds.mynetgear.com provides a robust framework for collecting game data, structuring it into an experience file, and leveraging that file during dedicated Learning Sessions to optimize engine performance.

The creation of an experience file represents a significant advancement in the field of chess engine development. By meticulously documenting and analyzing past games, developers can build engines that not only calculate moves with precision but also adapt intelligently based on historical performance. This approach is essential in today’s competitive environment where artificial intelligence is continually evolving, and where the ability to learn from experience is a critical differentiator.

The methods discussed in this article provide a clear and comprehensive guide for anyone interested in enhancing a UCI chess engine through the integration of experience files. Whether you are a hobbyist programmer or a professional in the field, following these best practices will help you create a robust learning system that continuously evolves and adapts to new challenges.

By implementing the strategies detailed in this guide, you will be well-equipped to develop a UCI chess engine that not only performs at a high level but also continuously improves over time, making it a formidable opponent in any competitive setting.

This comprehensive article has provided an in-depth look at how to create and utilize an experience file in a UCI chess engine. By integrating experimental data, carefully managed experience files, iterative Learning Sessions, and continuous feedback loops, you can develop a chess engine that is both smart and adaptive. Whether you are working with engines like Eman, HypnoS, or BrainLearn, these methodologies offer a pathway to enhanced performance and strategic depth in chess AI development.

source