carldaman wrote: ↑Tue Jun 29, 2021 5:31 am
Pointless only when trying to get the highest Elo overall. Others may prefer to have the ability to create specific playing personalities, whether human-like, anti-human, or Tal-like play-to-win attacking styles that the HCE could allow by means of tunable parameters.
I see no reason to believe that HCE and tunable parameters are necessarily the best or only way to achieve such goals. You could optimize a neural network for something else than pure results. A simple approach that would be interesting to try is to not just score wins as 1 and losses as 0 during training, but to score stylish wins (where "stylish" means "winning by using the playing style I'm trying to produce") higher than plain wins. You would still need to be creative and do some hard work, but instead of hand-crafting evaluation features, you would have to work on quantifying the stylishness of a game. What exactly is a Tal-like attacking win, and how do you recognise it in computer code?
Future nets could achieve the desired goals at some point. There has been some progress, notably Dietrich Kappe's nets, but tuning and optimizing NNs can be a slow and daunting task, whereas HCE parameters can provide quicker and more convincing results, at least at this time, provided that the HCE itself is very strong.
A Tal-like game, and not necessarily a win, would feature a healthy contempt for the opponent, a certain disregard for material, while favoring space and mobility, development, and kingside attacking formations. These can be coded and then further optimized, as you would well know. The results can be striking and spectacular once everything is properly weighted.
CyberNezh is a private A-B engine based on these concepts and its games can be viewed at lichess (see link in my sig.)
carldaman wrote: ↑Wed Jun 30, 2021 11:06 pmFuture nets could achieve the desired goals at some point. There has been some progress, notably Dietrich Kappe's nets, but tuning and optimizing NNs can be a slow and daunting task, whereas HCE parameters can provide quicker and more convincing results, at least at this time, provided that the HCE itself is very strong.
A Tal-like game, and not necessarily a win, would feature a healthy contempt for the opponent, a certain disregard for material, while favoring space and mobility, development, and kingside attacking formations. These can be coded and then further optimized, as you would well know. The results can be striking and spectacular once everything is properly weighted.
CyberNezh is a private A-B engine based on these concepts and its games can be viewed at lichess (see link in my sig.)
Good post!
Even if somebody develops a kit or a process to make the training of an NN to play in the style you like a very easy process, if you want a program to be able to offer the feature "specify the style of play", from what Carldaman wrote in the quoted post, HCE may well be the better way to go.
Now that a quarter of a century has passed, is Chris W willing to reveal how he got CS-Tal (link) to play such an exciting game of chess? Or maybe he already has?
Human chess is partly about tactics and strategy, but mostly about memory
i am testing CSTAL dos in the moment on fossil hardware (286-16 mhz).
its funny to see how it can beat other chess engines with doing only 300-450 NPS.
I do not believe that HCE is finished.
But i regrettably see that people have given up the idea to create an evaluation that could lead to something better.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....
mclane wrote: ↑Thu Jul 01, 2021 1:38 pm
i am testing CSTAL dos in the moment on fossil hardware (286-16 mhz).
its funny to see how it can beat other chess engines with doing only 300-450 NPS.
Sounds as though it was the Genesis era equivalent of LC0!
You worked with Chris on CS-Tal: can you give us any clues as to how it managed to play such an exciting game?
Human chess is partly about tactics and strategy, but mostly about memory
The elo chase killed HCE development. E.g., From my testing at different depths (10-20), I doubt that Komodo's evaluation improved from Komodo 8 to Komodo 12. It was clear that coding more knowledge would have to be a difficult task, and would have to be oriented to more specific knowledge, while most new programmers were oriented toward speed improvements and engineering, not chess specific tasks. When it was time to code more knowledge, NN were an easy --- and seemingly good --- solution.
In fact, this was the most glaring difference between the older generation of program authors and the new engine creators. 30, 20 years ago, chess was the focus of writing chess programs. Learning programming is the focus of writing an engine today. Most authors, engines are epigones. The few really original among the actives ones are either a bit older or chess players: Jonathan Kreuzer (who developed a beautiful interface), Jonathan Rosenthal (the first one to efficiently use NN in the evaluation), John Stanback, LK and a few others. There is less and less chess in chessprogramming, and it becomes glaring.
I think the clue is that chris, as a good chess player, was able to create evaluation functions that measure certain chess vocabulary.
And that he combined these functions with each other.
Instead of reducing the branching factor CSTal tries to find interesting situations and increases searching into these areas.
But this method only works if you have enough static knowledge about chess attacks.
Because you have to differenciate between positions that are good for the attacker and those not.
Otherwise your search tree explodes and you cannot see anything anymore.
By dragging the opponent (who has no knowledge) into these dark areas CSTal can fish into the fog and kill you.
That is what the tal function is doing,
CSTal is not searching each combination until the end and proofs if it works.
Instead relies on the evaluation function and plays the move, not knowing the outcome.
This is why the sacs can be unsound.
Speculative sacs.
For a human, such a playing is natural and normal. But for a chess engine it is unique.
Todays chess engines rely on search, nothing has changed this,
Search has made much progress. Evaluation is still something where you could do much.
But people are satisfied if they, by using search tricks, can make 200 elo. Then they lay back on the couch and think: why do i need to program complex evaluation functions when a few search ideas can do it too.
IMO this is like buying a pc trice as fast gives 50 elo.
Instead of changing the source code you buy a faster hardware.
That gives the elo that satisfies you and you dont have to investigate in the HOW TO DO.
This is IMO typical for our time.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....
Let me offer a contrarian viewpoint. First a few observations:
1. Most NNUE’s use an HCE as their starting point.
2. NNUE’s do best training on data produced from relatively shallow searches.
3. HCE’s are still used for hybrid searches.
4. HCE’s are optimized for AB search performance, not knowledge.
HCE’s could be optimized for hybrid search performance, e.g. you don’t have to be quite so sophisticated when you’re a rook up.
They could also be improved as the starting point for data generation. I’m sure every engine author has a few eval terms that they discarded because they were too expensive to compute. Perhaps it is time to dust those off and see if they help produce a stronger net.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
For chess you don’t need gigantic hardware.
But you need a plan. Or intelligent methods.
Thats IMO the quintessence.
If you have software appropriate and tiny hardware thats enough
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....
