FireDragon761138 wrote: ↑Thu Jan 29, 2026 8:06 am
Tactics are a big part of chess, obviously, and an engine would have no utility at all if it were weak at tactics. But the tactics should flow from positional advantages, to be humanly relevant in a generalisable way. The reason positions are often superior in chess is because they have latent potential to give rise to tactics. Think of knights in the center vs. the edge of the board, for instance. That latent potential can play into longe term advantages through prophylaxis and initiative far more than focusing on merely having complex concrete lines.
The focus on positional truth is about conceiving of chess in terms of narrative flow, which used to be an important part of chess pedagogy. Narrative is an important part of how humans acquire knowledge that integrates, and its more powerful for learning than simply teaching chess as an optimization problem.
Then I still think MCTS better represents your philosophy. MCTS operates more like the "expectation/potential" you speak of, while AB requires concrete lines to "prove" a position is good.
I have thought of that. I think optimism might help bridge the gap, it tends to help alpha-beta overcome its inherent pessimistic evaluation. And in my Theoria project, i'm not discounting concrete lines altogether, I'm just contextualizing them within a larger arc, having the position open to possibility rather than premature foreclosure. Humans in chess tend to dwell in the realm of possibility, if it were inevitable that one side or the other would win, games would be far less interesting.
An interesting idea might be to make a Stockfish fork, or similar engine, that uses a graphics card for a large neural network, ideally transformer based to capture some of the nonlinearities and holism of a chess position. I don't have the resources to do so at the moment, though. It would require getting more programming skill.
FireDragon761138 wrote: ↑Thu Jan 29, 2026 8:15 am
An interesting idea might be to make a Stockfish fork, or similar engine, that uses a graphics card for a large neural network, ideally transformer based to capture some of the nonlinearities and holism of a chess position. I don't have the resources to do so at the moment, though. It would require getting more programming skill.
I'm not sure if you're aware, but Lc0 has already been using transformer-based models since 2022.
FireDragon761138 wrote: ↑Thu Jan 29, 2026 8:15 am
An interesting idea might be to make a Stockfish fork, or similar engine, that uses a graphics card for a large neural network, ideally transformer based to capture some of the nonlinearities and holism of a chess position. I don't have the resources to do so at the moment, though. It would require getting more programming skill.
I'm not sure if you're aware, but Lc0 has already been using transformer-based models since 2022.
That's true, but Lc0 doesn't fit within the design parameters I'm looking for- computational efficiency. Otherwise I probably would use something like Lc0. Stockfish's underlying architecture is just less fussy than Lc0's in general, too, especilaly when you get to the UCI options. I view Lc0 as more a testing platform, in that respect, than a practical tool.
Well, if one could come up with a neural network architecture or search method that is more efficient than Lc0's, that would be a genuine breakthrough that would get a lot of interest from the Lc0 team.
sscg13 wrote: ↑Thu Jan 29, 2026 8:34 am
Well, if one could come up with a neural network architecture or search method that is more efficient than Lc0's, that would be a genuine breakthrough that would get a lot of interest from the Lc0 team.
That's what more or less what I'm trying to do with Theoria. It doesn't have Lc0's raw power, but it does have more efficiency (I might even experiment with bigger networks in the future, as initial tests look like that might be promising, even for an NNUE, for my purposes).
I'm going to do more testing on Lc0 at high level of play (maybe rent some time on Runpod, or even run Lc0 on my own machine if I have to), to compare that to both Stockfish and Theoria and see if there's any correlations or analogies between the patterns in the data. And that might clue us into some answers about what optimal play might look like, beyond speculation, by reasoning by analogy.
FireDragon761138 wrote: ↑Thu Jan 29, 2026 5:57 am
The latest study I did showed that we can achieve several orders of magnitude more computational efficiency in the search, simply by changing the evaluation training regimen. That means it's more parsimonious and energy efficient. You don't get 22+ moves to achieve engine evaluation stability, and the resulting evaluation is still above what is needed for the typical human chess player to analyze their games.
I know something better: just limit SF to 1ms per move.
Now run your "study" again (i.e. ask DeepSeek).
Voila.
OP, please decide if you refer to yourself as "I" or "we", it's a bit inconsistent. Please discuss it with your friends and ghost writers Chad and Claude.