One week ago, I left TCEC Season 9 Superfinal Week 2 with Stockfish leading Houdini by a score of +7-2=31. During the intervening week another 29 games have been played with a cumulative score of +5-3=21, again in favor of Stockfish. That makes an overall score of
+ 7-2=31 after week two
+ 5-3=21 week three
Five wins for Houdini is a better score than I had extrapolated after the first few days of play. How does it manage to prevail against an opponent who consistently calculates variations more deeply? The following chart shows the evaluation during the course of the game for four of the five wins, Houdini playing White each time.
The beginning of each game shows a similar pattern. After the mandatory book moves have been played, Houdini gives a higher evaluation to the position than does Stockfish. After more moves have been played, Stockfish's evaluation eventually catches up to Houdini's, then surpasses it as Stockfish realizes the game is lost.
Is Houdini's evaluation of an opening position more accurate than Stockfish's? To help answer that question, I would compare the results of the same openings with Stockfish playing White, but that will have to wait for another time. It's worth noting that Stockfish also won G65, i.e. playing the White side of G66. Perhaps that provides an additional clue to its evaluation of openings.
Houdini won another game not shown here, game 58 (G58). The opening of that game followed the same pattern as the other four games, but something strange happened in the middlegame. At around move 65, in a position where the center was blocked, Houdini (White) calculated that it had an advantage of a third of a Pawn, while Stockfish evaluated the position at 0.00 (dead even). The engines maintained their evaluations for another 30 moves, after which Houdini's dropped to a slight advantage for its opponent. After another 15 moves, Houdini's evaluation climbed to a half-Pawn advantage for itself, then continued to climb with Stockfish still showing 0.00. A few moves later, both engines gave a winning advantage for White. On move 144, the game was declared a win for Houdini.
It's well known that engines often have a problem evaluating blocked positions. Was this a contributing factor to the other four wins for Houdini?
While I was writing this, Houdini also won game 70. Although it's putting up a better fight than I had anticipated, I don't expect the overall match result will change in the 30 games left to be played. At a rate of four games per day, we'll find out in another week.