Tech mate

I am sure I have posed more questions to the chess engine Stockfish than to any living being. I love the instant gratification: you give it a chess position and it gives you an answer: the best move and an evaluation measured in hundredths of a pawn, like +1 24. Leave it alone, it will delve deeper; move a piece, it will respond anew. In human terms, ‘Ooh, a smidgen better for White’ gives way to ‘Go there, and your rook gets blown away’. The chess engine is, by turns, a spirit level and a hurricane forecast.

The basic design of a traditional chess engine like Stockfish is simple: check all the moves, back and forth, for both sides, and tot up the value of the pieces in each scenario. Then choose a move accordingly. But the tree of possibilities is so vast, even for a computer, that you must prune it judiciously: ignore the moves that blunder a queen, but not the brilliant sacrifices that lead to checkmate. Evaluating the pieces is nuanced as well. In the middlegame, you want mighty centre pawns, but in the endgame, a nimble flank pawn could be better. And don’t be too clever! If the evaluations are too sophisticated, they will take more nanoseconds to compute, so you won’t be able to examine as many positions.

Deep Blue, which narrowly beat Garry Kasparov in 1997, is a minnow next to the Stockfish of 2020, which is not just down to improvements in hardware. The powerful blend of algorithms and parameters that comprise Stockfish has been long in the making. An initial version was derived from a program called Glaurung in 2008, written by Tord Romstad. The latest version, Stockfish 11, credits Marco Costalba, Joona Kiiski and Gary Linscott alongside him.