Interview

“Leela offers us a second opinion”

26 Apr 2020

Matthew Sadler is the leading authority on what chess players can learn from artificial intelligence and neuronal network based programmes. After Leela beat the strongest traditional chess engine Stockfish in the final of the 17th TCEC chess computer super league season, he shared his conclusions with Conrad Schormann. Even though the neural networks now seem in the lead, the clash is far from over and is promising to produce more fresh chess ideas.

Matthew Sadler and Natasha Regan presented Game Changer at the London Chess Conference 2019
Matthew Sadler and Natasha Regan presented Game Changer at the London Chess Conference 2019 (photo: John Saunders)

Matthew Sadler is an English grandmaster and consultant specializing in IT architectures and infrastructures. He retired from professional chess when he was 25 and a top 50 player. Together with Natasha Regan he wrote Game Changer: AlphaZero’s Groundbreaking Chess Strategies and the Promise of AI (New in Chess 2019). Follow Matthew Sadler on Twitter.

Matthew Sadler’s YouTube Channel

The superfinal between Leela and Stockfish is over. You followed the 100 games closely and tweeted frequently. What’s your take on the match?

Matthew Sadler: The contrast in styles, the contrast of the positions in which they excel, how they are both extremely strong while managing to expose each other’s weaknesses – fantastic. There were some games in which one engine was killing it while the other looked helpless, and that happened to both sides. This time Stockfish was absolutely lethal in open Sicilians, essentially in every open position in which king safety is a factor. On the other hand, the range of positions in which Leela can actually hurt Stockfish has become somewhat greater. Leela did very well in King’s Indian and French type structures, essentially in any position with a locked center and play on both wings.

The more concrete calculation is required, the more Stockfish shines. Is that also true the other way around?

Matthew Sadler: In all fairness, Leela’s calculation is stunning as well. But, same with AlphaZero, decisions about the course of action are taken quite early, and there’s always the chance for the neural network to miss an unusual move or a tactic early in the calculation tree. But when calculating deeply it is amazing what Leela can do.

“Playing through AlphaZero’s games for the first time was one of the big emotional events of my career. Suddenly I realized how much more there is to chess.”

52.5 to 47.5 in Leela’s favour (+17 -12 =71) may sound like a clear result. Peter Heine Nielsen tweeted that this is very much within the margin to be expected between two equals. How do you see it?

Matthew Sadler: The race is still on, we will enjoy more rivalry. During the match the Stockfish developers announced lots of patches, lots of new stuff coming. To me it seems still unclear which approach is the best for chess. Take Go for example, played on 19×19 squares with multiple conflicts all over the board: there is just too much going on to calculate it. The neural networks with their so-called intuition handle it much better. Chess on the other hand, 8×8 squares, lots to calculate, that’s a game where you can’t write off Stockfish. And, frankly, that’s great! We will enjoy more matches with the contenders becoming better and better. And we will see that there is still a lot to discover in chess.

Who would have thought …?

Matthew Sadler: Exactly. When I retired as a professional player in 1999 the engines were already considered very strong. I thought, chess may be finished in a couple of years. Well, here we are! Twenty years later it’s still a fantastic game, rich, unbelievably complex, and we have two great engines, neither of them perfect. This development is a testimony to the richness of our game.

Matthew Sadler and DeepMind CEO Demis Hassabis
Matthew Sadler and DeepMind CEO Demis Hassabis (photo: Lennart Ootes/London Chess Classic)
How would your beloved AlphaZero do against today’s Stockfish and Leela?

Matthew Sadler: Difficult to say, especially for me. One of the big emotional events of my chess career was when I played through the AlphaZero games for the first time in 2018. DeepMind had asked me if I’d be interested in doing this. Of course, I expected something strong. It’s DeepMind after all, these guys are serious, so I knew it would be quite something. Still, AlphaZero’s style hit me completely unexpected. At that time, I considered Stockfish to be pretty much optimal. Suddenly I realized how much more there is to chess. Wow! So I’ve got this emotional connection to AlphaZero, this feeling that surely nothing could ever be better. On the other hand, as an IT guy I know that stuff needs to be updated. Otherwise, however magnificent it was, it gets overtaken. And it’s been two years almost. Stockfish and Leela must have come very, very close.

How to make them even better?

Matthew Sadler: With Stockfish you can tinker directly, it is made by humans after all. Improving the search, let it look deeper and further is an ongoing process. But there is also knowledge to implant to help it handle positions that it struggles with, all sorts of things. There is for instance a Stockfish clone specialized in understanding fortresses, something the regular Stockfish has trouble with. With the neural networks on the other hand tactics and time management are fields that are open for improvement. But you can’t just tell it to do something differently, Stockfish has an advantage in this respect. Leela needs to be given the conditions under which it can learn as much as possible. That is a lovely contrast there as well.

What can human chess players take from the recent match in order to learn from it?

Matthew Sadler: Every professional needs to look at these games. Most professionals use Stockfish for analysis in order to filter out every tactical mistake faultlessly. Now it’s interesting to see Stockfish faltering – and that may help people’s preparation. Leela is strong enough to expose where Stockfish is getting it wrong. I can say from experience that before Leela or AlphaZero I have played lines even if they looked dodgy to me just because Stockfish said they are ok. Now Leela offers a second opinion on these on a more conceptual level. That is an important aspect for professionals. However, most opening lines in the match weren’t quite mainline, but still there is remarkable opening stuff to study. Leela’s handling of the Mar del Plata Variation of the King’s Indian for instance. Also, the King’s Indian Sämisch games were interesting from a theoretical point of view.

“Before Leela or AlphaZero I have played lines even if they looked dodgy to me just because Stockfish said they are ok. Now Leela offers a second opinion on these on a more conceptual level. That is an important aspect for professionals.”

Any advice on how to use the different strengths of Leela and Stockfish for human opening preparation?

Matthew Sadler: One of my and some other people’s favorite way of opening preparation is to come up with interesting opening positions and have Leela and Stockfish play 100 game matches from these. The games that you get from that are a huge help – especially when learning new openings. You not only get all the evaluations; you also get a body of experience that is only yours. And it’s a low-cost way of getting stuff done: I think about an interesting position for a few minutes, try to come up with ideas, then I press a button and say “go on, boys!”. The next day I take an hour to go through the games, take out the best ones. These I put into a file that will be filled with inspiration and gems I can draw from. That’s something we couldn’t do before the neural networks came along because traditional engines were so similar.

With “Game Changer” you and Natasha have written the most important chess strategy book in recent years. How about “Game Changer II”?

Matthew Sadler: With writing books, it always takes me a couple of years to forget how much effort that was before being brave or foolish enough to start again. Sure, it was a fantastic project …

… and valuable! We want more. Someone needs to break down what is happening.

Matthew Sadler: Yes, that was the nice aspect of Game Changer: to make this super level chess accessible to the club player. And there still are lots and lots of material. We already could have done a book three times the size of the actual one, now that we have all the new developments. I do have many ideas what I could write about. I am always thinking, but there is no current project to talk about. You never know, it might happen. In the meantime, I love doing YouTube.