Zenmastur wrote: ↑Wed Jul 10, 2019 3:05 pmA question. What is the best hybrid system right now in your opinion.
It's called Eman XXX. I have no idea about how it works, as apparently it's some private thing owned by a few individuals. But Leela and Stockfish (Eman is strongest derivative) are run in parallel and the system does something to pick a move.
My guess is that there's some positions where Leela is worse than Stockfish and others where Stockfish is worse than Leela, and the system is picking moves from the worse entity, which proves fatal, or something.
So the hybrid performs worse than Stockfish or Leela alone.
I know it's the best hybrid because these people want to win very badly and are using the best that they can, if a better hybrid existed they'd be using that instead. So it turns out we're at a point where the computer plays stronger with the CPU sitting idle than with the strongest A/B engine chunking out moves and scores that just weaken Leela, and Leela is improving faster than hybrid efforts, so unless Leela hits a brick wall at some point (flat lines on her improvements), gambling for hybrid looks very risky.
Zenmastur wrote: ↑Wed Jul 10, 2019 3:05 pm My idea for a hybrid system would be something like training a small NN to do the LMR for a AB engine ( this seems to be where SF seems to have a weakness). Same for which TT entries to evict during replacement operations. Integrating a self-learning book that selects lines of play based an engine performance in those lines, not necessarily using NN for this one since there is a way to do it that uses less intensive calculations methods.
Right, looks like hybrid is still on its infancy and that a good approach would need to do something different at the learning phase of the NN.
Zenmastur wrote: ↑Wed Jul 10, 2019 3:05 pmI have long been of the opinion that iterative deepening, while it works very well, is a sub-optimal search strategy once a certain fraction of the available move time is used. The exponential growth in search time with search depth seems to bear this out. This also inhibits plan finding in AB engines because in too many cases no useful plan can be found with the limited search depths. Often times the plans that are found are sub-optimal again as a result of limited search depth. A mixed search strategy based on shorter compound ID searches seems like a reasonable approach.
Yeah, I can't help but think A/B engines are wasting so many resources and that a better multicore strategy must be devised. When I take a look at people with 40 cores playing 10 ply deeper than me I just think I should be losing the game badly, so drawing with ease just means A/B engines are using the extra resources improperly.
I would note that just like AB progams NN programs could greatly benefit from a self learned opening book. This would free up more time once out of book and could change the effectiveness a particular size NN.
Currently self-learning books have been tried for Eman (with its experience file) and SugarNN (where people start without book and let the thing improve and learn by itself) and for the former Khalid (Eman's author) abandoned it in favor of normal books (the engine uses experience file only after leaving book) and SugarNN is playing weaker than Eman, but I guess it'll improve as time goes by...
I think the break point is when you analyze the games after and you can't find any mistakes using an AB engine. When this starts happening on a regular basis then I would be worried.
It's already happening on a regular basis with A/B engine Vs. A/B engine, regardless of hardware (unlike 5 years ago, where a hardware like mine was suicide, and this was against top hardware of that time, which is 5 year old hardware by now). It's still happening rarely with A/B Vs. Leela. Looks like Leela's success is arriving at positions where A/B engines are much more likely to blunder, regardless of hardware (A/B engine being "outevaluated" so it doesn't matter what depth they reach), but where Leela is also likely to blunder (again, regardless of hardware, so if the position requires high tactical capability, the strongest piled GPUs will lose against A/B in hardware slower than mine.)
You could, but who would want to? It's way too tedious with slow hardware. In CC chess you are likely to be playing 50 games at a time (or more), with time controls of 10 moves in 50 days. This mean your average mover rate must be 10 moves per day. If you work, sleep and have a “normal” life this leaves very little time for the analysis. This is why you run your computer 24/7 365. You clearly get much better analysis if you are running 25M nps with a 64GB TT than if running 2.5M nps with an 8GB TT.
In addition when sitting at the machines your interactive analysis goes much faster, so you get more done.
Diminishing returns and self-refutation (where after enough analysis you find the best move and must play it, because more analysis leads to refuting it - even if the opponent wouldn't - and second best move "catching up" so you no longer know what move is best) are a huge problem in correspondence games, so that instead of playing your games at a little bit higher level (50/10 time control is long enough that you have enough time to reach the breaking point) you're better off just playing more games.
Upgrading hardware to be able to play 80 games instead of 50 doesn't make sense to me, because there's no benefit in playing more games. The branching factor of chess just doesn't make upgrading cost-effective after you reach a certain point, and I believe Uri is already beyond this point.
Your beliefs create your reality, so be careful what you wish for.
