---

An open-source AI chess engine

Leela Chess Zero (LcO) is a great project. Gary Linscott has implemented a neural-network (TensorFlow) based chess engine that runs on generic hardware. Like AlphaZero (AZ) it developed its skills entirely through self-play starting with random moves. Thanks to the chess community, self-training is running at around half a million games a day and the evidence is that Leela has achieved grandmaster strength. You can run the training software on your own computer and contribute to Leela's chess intelligence: it's very straightforward (got it running here right now) just visit the Github link below and follow the instructions.

Why is Leela interesting? AZ led the way and played mind-bogglingly good chess in a human, romantic style that felt very different to a strong algorithmic program like Stockfish. However, AZ is locked to Google's expensive custom hardware and Joe Public can't play against it—it's a closed project. Leela is an open project: you can participate in her progress; and you can play against her!

Play against Leela Stats and Info On Github

Will Leela be as strong as AZ was?

The maths suggests she will—assuming her competitive games are on equivalent hardware to AZ! Deepmind's hardware trained many thousands of times faster than possible using consumer CPUs or GPUs (estimated 1,700 years worth in 4 hours), however the distributed training model will help level things! Artificial Neural Networks (ANNs) all solve tasks in similar ways, and it is accepted that they can duplicate any mathematical function. So given enough time and data, all the subtle input-patterns that affect the game will become incorporated in the synapse values of a suitably configured ANN. We know the underlying methodology behind AZ, so without any secrets, Leela should catch up.

There's a buzz about Leela just now in the chess community. She is registered with Lichess as a bot, so has established a public persona in the chess world. We'll be hearing a lot more about Leela.

I Do have Worries, though.

Chess ANN training using wins in randomised play has, it seems to me, one basic problem. Ultimately it could produce an entity which plays the best winning chess on the planet, but it may simply end up as a better Stockfish arrived at by a different route. Personally, I don't want to play a perfect super-grandmaster, but a bot which plays good chess AND has a sense of fun, inventiveness, and experimentation. That does not mean adding random blunders, or even random lower-ranked moves, but a whole mindset where the software will spring an interesting surprise just to see what happens, and have a follow-up attitude which depends on the opponent's response.

One of the joys of early Leela was that the way she played felt very human, now her style feels more like being slowly crushed in the coils of a boa-constrictor. We know ANNs can play grandmaster chess, now we need some serious effort put into the human element because it will not happen automatically just because we are using an ANN.