Komodo MCTS: the boa constrictor approach

Testing Monte Carlo Tree Search

When Deep Mind unveiled their AlphaZero, it broke the chains of status quo irrevocably on a variety of fronts. On the one hand, there was their self-learning neural network, and on the other, there was their use of the MCTS search system, which is vastly different from the hitherto ubiquitous alpha-beta search system that ruled the day.

Because of how new it is, very little advanced development has been made to fine-tune it, and as a result, it still performs worse on a pure Elo level. However, Mark Lefler, the chief programmer of Komodo, has been working hard with nearly daily breakthroughs, and promises to release updates soon (“but [we] are waiting for the rather significant daily Elo gains to slow a bit.”), available for free to owners of Komodo 12.

That said, what is the value of this new MCTS system, and does it work even when not coupled to a neural network? The answer is a very vocal ‘yes’.

Have you ever watched a game of Magnus Carlsen, as he does his unique boa constrictor act on his opponents, and wondered how he seemed to be able to win these ‘drawn’ endgames? Perhaps you are watching with one of the top engines, and it invariably says the position is dead drawn until all of a sudden with the snap of a finger Magnus is winning. Is it Voodoo? Hypnotism? Or was the engine just wrong? It is actually none of these things.

What a player such as Carlsen does is to outplay his opponents, even though the evaluation by the engines is a big fat 0.00. How can he be outplaying them if the evaluation never changes? Behind the curtains of that game, he is slowly reducing the number of available moves to his opponent. At first, they might have three or four perfectly viable choices, all of which draw, but as the game proceeds, they now only have two or three viable choices, until before long they are walking a tightrope act with only one move to keep it together or lose.

For classic engines, as long as there is one single move that can save, they will declare it a draw. Objectively, they are not wrong, since that one saving move does save, and it is there, but that doesn’t begin to tell the story. MCTS search looks at the various options, some losing, some winning, and some drawn, and chooses the move that offers the best chances to win or save. Naturally, if there is only one right move, it must choose it, but if given a fork in the road, then much like Carlsen, it will choose the continuation that presents the most practical chances.

In order to illustrate this, GM Karsten Mueller has shared a few examples that show the difference, and value, of this sort of analysis:

Tricky to the end — Part 1

In the next example, we see the MCTS search play like a grandmaster, not because of 'perfect' moves, but for making it as hard for the opponent as possible to keep control.

Tricky to the end — Part 2

In the following position, White exploits some tactics to give Black more trouble than he had imagined.

Unexpected troubles

As the grandmaster explains, it is a mistake to just push one's pawns blindly forward in an endgame with passed pawn by both sides.

When the knight trumps the bishop

Chess Endgames 14 - The golden guidelines of endgame play

Rules of thumb are the key to everything when you are having to set the correct course in a complex endgame. In this final DVD of his series on the endgame, our endgame specialist introduces you to the most important of these rules of thumb.

In the next and final example, the grandmaster cannot help himself as his teaching instincts take over and he serves up a highly illustrative example of Rauser's Drawing zone. Most readers will be familiar with the concept of the h-pawn and the wrong bishop to ensure a win, but there are exceptions, as below:

As GM Mueller explains: "Black wins, if he can force White's king to leave the zone illustrated by the green squares, permanently."

Almost counterintuitively, White played 98.Kb4?? and lost. See the game to know how and why.

The exception confirms the rule

As you can see, the MCTS search has valuable insights to offer by its very own nature. Hopefully it will also help you make the most of your resources.

Correction June 19th: GM Dr Karsten Müller agrees that the definition as originally given above for "Rauser's drawing zone" was not correct, and it has been amended.

Komodo 12

In computer chess there is no getting past Komodo, a two-time ICGA Computer World Chess Champion. Find out how Komodo can take your game to the next level!

Links

• Komodo 12 with AlphaZero techniques
• The future is here – AlphaZero learns chess  
• Alpha Zero: Comparing "Orangutans and Apples"