Man vs. Machine, Parts I+II
In life as in chess, we are at the mercy of machines.

Man vs. Machine, Parts I+II

the_real_greco
the_real_greco
|
15

Deep Blue vs. Kasparov. Deep Junior at Dortmund. Deep Fritz vs Kramnik. Leelenstein vs jjosh?

The history of humans playing computers is not a long one. The period in which chess engines and strong human players could compete on level footing was 10 or 15 years long, at most. Before that time computer chess was something of a punchline; currently, it is accepted that a top-tier engine on reasonable hardware can easily crush the strongest human players.

But there is nothing like a good human-engine match to entertain spectators. For that reason I present here a collection of engine developers losing to their creations. These engines are stronger than anything Kasparov or Kramnik ever played... and their creators not exactly grandmasters. 


Is it Leelenstein, or Leelenstein's Monster?

First up is Leelenstein, a neural-network engine custom-trained on millions of high-level engine games and correspondence chess games. It began with the same binary as Lc0, but is being rapidly developed with unique updates. Leelenstein placed third in CCC8, behind only Stockfish and Lc0. Leelenstein's mysterious creator is known only as jjosh.


She might look innocent, but even on one node, she plays a mean game of chess!
Next up is Allie, a neural-network engine with original binary. It makes use of the same network as Leelenstein. Her creator, Adam Treat, is known to the CCC community as gonzochess. In the future he hopes to have her using alpha-beta (AB) pruning, rather than her current Monte-Carlo Tree Search (MCTS). In this game Allie is severely hampered, using no search but instead relying on pure intuition. (Spoiler: She won anyway).

Most snowflakes are hexagonal; this one is not.
Next up is Winter, a new engine by creator Jonathan Rosenthal. It was created in 2016 as a school project (fun project, right?). Winter is an AB engine with a handwritten evaluation function, although it also uses traditional (non-NN) machine-learning techniques. Version 0.6 is currently competing in CCC9. 


ExaThought, known for its extremely unique style.
Last up is ExaThought, named after it's creator Exa. Have you ever wondered how an engine would play with a primitive evaluation function (piece values and pawn-advancement bonuses) and no move pruning? ExaThought is the engine for you! In this game it uses a fixed search depth of 7 ply, meaning it moves its queen around- a lot. While that is enough to beat its creator, ExaThought might be a few years away from the CCC.

The last game is me against Stockfish 9, included as a promise to the engine developers. Misery truly loves company.