Antecedent
No prior reading is required. However, this article is intended for readers familiar with chess.
Emergence
Skimming through chess opening theory, I realized that many opening sequences, even those people barely use, have been named after someone or some place. This implies that if you play several strong moves in a row, you have likely, perhaps by accident, played an existing, named opening.
Therefore, I wondered: can I identify legitimate opening sequences that have not been named yet?
Stabilization
To "create" an opening, I first needed to set a precise goal. The goal isn't just to make several random, nonsensical moves and name them. It is about finding hidden, unnamed, and legitimate opening sequences. If I simply name an awkward sequence of moves that nobody plays, it's meaningless.
So, what defines a "legitimate" opening?
A legitimate opening, in this context, means that both players have consistently played the best (or near-best) moves available. To measure this objectively, I created the following cost function: are respectively the total opening with moves, the state after th move and the evaluation function)
This function measures "how bad the worst move in the entire sequence was," based on the change in evaluation score. An opening is considered more "legitimate" as this cost function gets lower.
Now, the search process can be automated. Stockfish is a powerful tool for evaluating a given position. I wrote a program that uses a beam search (which involves keeping only the top-ranked moves at every stage and expanding from them) to find sequences with the lowest cost, up to a fixed total length.
The program then sorts these sequences by their cost. Finally, we can go through the sorted list, starting from the one with the lowest cost, and check each sequence against the Encyclopedia of Chess Openings (ECO) to see if it is already included and named.
This is the code.
There was a slight problem in this process. Stockfish's evaluation, even at a practical depth, tended to value quiet openings (like starting with a knight) more highly than their common reputation might suggest.
Thus, a lot of quiet and seemingly boring sequences ranked higher than openings we frequently use, like the Ruy Lopez and the Queen's Gambit. So, we can't conclude that this sorted order is perfect. Instead, we must use it as a guideline to find new openings.
When I set the total length to 1 (i.e., a single half-move), we get a trivial result. Every such sequence is already named, so our algorithm simply ordered them.
Here are the first few best sequences that yielded the lowest cost function values:
- 1. Nf3: Zukertort Opening (A04)
- 1. e4: King's Pawn Game (B00)
- 1. c4: English Opening (A10)
- 1. d4: Queen's Pawn Game (A40)
- 1. g3: Hungarian Opening (A00)
When I extended the length to 2, the results were similar.
- 1. e4 e6: French Defense (C00)
- 1. e4 e5: King's Pawn Game (C20)
- 1. e4 c5: Sicilian Defense (B20)
- 1. e4 c6: Caro-Kann Defense (B10)
- 1. d4 Nf6: Indian Defense (A45)
But... lower down in the ranks, there was quiet and unknown move hiding.
- 1. d4 c6: Unknown Opening (N/A)
But this is too short to be considered a specific opening. Let's extend the search to the next move.
After extending the search depth further, things got strange. Stockfish highly rated most of the quiet moves involving knights, placing them all over the leaderboard. This led to unproductive results. Therefore, I needed to artificially restrict the first move to either e4 or d4. Here are the results from that filtered search.
- 1. d4 Nf6 2. Nf3: Indian Defense: Knights Variation (A46)
- 1. d4 Nf6 2. c4: Indian Defense: Normal Variation (A50)
- 1. d4 d5 2. Nf3: Queen's Pawn Game: Zukertort Variation (D02)
- 1. d4 d5 2. c4: Queen's Gambit (D06)
- 1. e4 c6 2. Nf3: Unknown Opening (N/A)
- 1. e4 c6 2. Nc3: Caro-Kann Defense (B10)
The fifth sequence in this list looks like the first known opening of this length to appear in our top rankings. However, it's just one of the main lines of the Caro-Kann Defense that is not formally included in the ECO.
I think this is the highest-ranking truly unknown opening found by the algorithm, which was ranked ninth.
- 1. d4 c6 2. Bf4: Unknown Opening (N/A)
In the same way, I discovered what I consider the best unknown openings (candidates to be named) at total lengths of 4.
- 1. d4 e6 2. Nf3 c5: Unknown Opening (N/A)
Convergence
Now that new, unnamed, and legitimate openings have been discovered, only one task remains: the naming process. After long reflection, I have chosen a name for these openings.
I will call them... the "Unknown Opening."
Descendant
No more related particles, yet.
| |
|---|---|
Phonon | |
Phonon | |
Tachyon | |
Phonon | |
Gluon | |
Tachyon | |
Phonon | |
Lepton | |