Chess AI engine in 5 years.

[Mike Sherwin]

You have had the opportunity to actually try to implement since the advent of NNUE for several years now and it isn't our fault that you limit yourself to talking about it on this forum.

Your ignorance of my situation is not your fault. What is your fault is judging me. I was born with a learning disability. That learning disability has only gotten worse with age. I barely was able to code RomiChess in the first place. I have not been able to bring a new chess engine to completion since. Plenty of people have tested Romi's learning in the past. You do not have to just take my word for it. But go ahead and tear me down some more if it makes you feel superior!

[Mike Sherwin]

Well let's be real no one is there to test smth extremely vague spitted out on the forum to see if it gains.
I have a pretty big experience in developing my own ideas as well as implementing other people ideas but in general the best way would be to DYI, after all it's not that hard to try stuff for sf if you are a capable programmer, also it really kills any need in you having hardware to test it since fishtest provides it for everyone.
What an opportunity for you, don't you think?

I am not a fully capable programmer. I never was. I was born with a learning disability. It has only gotten worse with age. Your ignorance of that fact is not your fault. Your words are your fault though.

[Mike Sherwin]

So let's get back on track. The original question was, "How miuch better will chess AI engines be in 5 years? How much more room for improvement is there".

I'm answering the second part of that question. And the answer is a lot better. There is a lot of room for improvement. While current NNUE is the best so far is has loads of possible suboptimal values for any given position. If I am right about pre search learning then +1000 elo would not be unreasonable.

[JacquesRW]

Your ignorance of my situation is not your fault. What is your fault is judging me. I was born with a learning disability. That learning disability has only gotten worse with age. I barely was able to code RomiChess in the first place. I have not been able to bring a new chess engine to completion since.

I'm sorry about that. It does not mean you can expect that if you make strong claims that no one will question it or provide pushback. I think I have been totally fair and have attempted to make valid criticism (as well as suggesting ways in which you could get your idea tested for you, even).

Plenty of people have tested Romi's learning in the past. You do not have to just take my word for it.

I'd much prefer to have actual reproducible evidence (which should be trivial to provide, I should just need a RomiChess binary?), rather than "trust me bro" from people I don't know who tested RomiChess years ago.

But go ahead and tear me down some more if it makes you feel superior!

I'm not saying this as a personal attack lol, I'm expressing scepticism and I have *tried* to establish a path forward for 1) validating your claims about RomiChess and 2) testing your idea in a modern engine.

If I am right about pre search learning then +1000 elo would not be unreasonable.

This amount of elo for top engines would certainly be unreasonable. Keeping your claims in the realm of somewhat believable would do a lot for people wanting to help out.

[Leo]

Thanks for everyone's comments. Very helpful. I am looking forward to ongoing improvements and refinements in the top chess engines with new hardware and improved input.

[Viz]

Thanks for everyone's comments. Very helpful. I am looking forward to ongoing improvements and refinements in the top chess engines with new hardware and improved input.

Most improvements for years were (apart from NNUE) and still will be improvements of search selectivity and move ordering.

[Mike Sherwin]

I'd much prefer to have actual reproducible evidence (which should be trivial to provide, I should just need a RomiChess binary?), rather than "trust me bro" from people I don't know who tested RomiChess years ago.

Not trivial. I had to resurrect my MediaFire account from the ashes of forgotten passwords and forgotten email accounts.
https://www.mediafire.com/file/ju22evcm ... n.zip/file
And as a bonus https://www.mediafire.com/file/yjei0ni5 ... k.zip/file if anyone likes simple but fun games.

[Dann Corbit]

Thanks for everyone's comments. Very helpful. I am looking forward to ongoing improvements and refinements in the top chess engines with new hardware and improved input.

Most improvements for years were (apart from NNUE) and still will be improvements of search selectivity and move ordering.

I think this is true wisdom. On the one hand, CPU power will eventually come to the point where LC0 sized nets can run on the CPU, but at the same time there will also be transparent access to the same memory fields for the CPU and GPU. I expect a lot of power to come from the network.

But revolution in strength comes from the search improvements. Not just reduction of the branch factor, but intelligent reduction. The big jaw dropping improvements in chess engine strength have come from search improvements (with the one exception being the introduction of using a network for eval, which was eye-popping initially, but now I think the progress is much more incremental.)

That move ordering would be so important is a new idea to me. Obviously, bad move ordering will destroy a search, but I would have thought that move ordering was close to ideal. There were some interesting ideas in Matthew Lai's paper on Giraffe. One of his main ideas was using AI not just for evaluation but also to guide the search. It seems that there may be more horsepower in that area to come.

[Viz]

Just recently I introduced new heuristic in move ordering - sure, it's not the biggest of gainers but it for sure shows that there is more to be done there, and mind you, stockfish has a really sophisticated move ordering (compared to other engines).
Also this idea works in Alexandria, for example, so not SF specific.
https://github.com/official-stockfish/S ... f61415b4e6
https://github.com/official-stockfish/S ... 3127f85509
https://github.com/PGG106/Alexandria/co ... 02409c4e05
In general leela with pure move ordering (aka just using policy) is GM+ strength, stockfish isn't even close to this, so there is definitely some a lot of space to gain.
Maybe at some point someone will come with efficiently updated NN for move ordering, for example, current one is handwritten almost fully (almost - because of this part using static evaluations of net to adjust move ordering, so it's automatically generating policy on the fly... Somewhat. https://github.com/Vizvezdenec/Stockfis ... h.cpp#L762)

[Viz]

Also about selectivity - at times of let's say stockfish 8 usual depth 13 bench run took like 8 million nodes, nowadays it's barely above 1 million and I think we actually added some positions since then.
And this kinda remains the same with higher depths - in fact with fixed depth search stockfish new releases are usually worse than the previous one, difference is that they reach this depths much faster.