TCEC S15, END of an ERA event is much more Brutal than I thought!

[M ANSARI]

Lc0 has improved on many of the weaknesses it had and is playing quite solid ... only problem seems that time management can be improved. Hard to understand why obvious and only move captures can take many minutes to play. I was watching online one of the games where Lc0 played an incredibly effective dynamic defense in what seemed like a losing position and held SF at bay to what seemed like a sure draw ... only to blunder as time dwindled down to seconds. According to the chat this time management issue caused earlier losses as well. It almost seems like Lc0 has to create a brand new memory hash table from scratch on every move and somehow does not keep in memory moves it has searched earlier. An obvious move capture that has no other possible alternative should not take several minutes to play and maybe that is something that can be easily patched. Does anyone know how Lc0 stores its hash and is it able to access this stored information on the next move?

[chrisw]

Lc0 has improved on many of the weaknesses it had and is playing quite solid ... only problem seems that time management can be improved. Hard to understand why obvious and only move captures can take many minutes to play. I was watching online one of the games where Lc0 played an incredibly effective dynamic defense in what seemed like a losing position and held SF at bay to what seemed like a sure draw ... only to blunder as time dwindled down to seconds. According to the chat this time management issue caused earlier losses as well. It almost seems like Lc0 has to create a brand new memory hash table from scratch on every move and somehow does not keep in memory moves it has searched earlier. An obvious move capture that has no other possible alternative should not take several minutes to play and maybe that is something that can be easily patched. Does anyone know how Lc0 stores its hash and is it able to access this stored information on the next move?

We discussed the missing of obvious move captures on Programers section a few weeks ago.

[Gary Internet]

The real end of an era is happening in engines that are TCEC Divisions 1 to 3. There don't appear to be any up and coming AB engines with the exception of Rofchade. The only AB engines to ever be stronger than Fire 7.1 are Stockfish, Houdini and Komodo. Ethereal was getting close and has now ceased development. Everything else close to Ethereal's level of strength appears to have ceased development as well.

Most of the TCEC league is now stagnant because AB engine development is waning.

Based on where the engines will start season 16, this is what the situation looks like:

Division 1
1. Fire (hopefully Fire 8 will be released later this year)
2. Ethereal (Nothing since 30 March 2019)
3. Xiphos (Nothing since 30 March 2019)
4. Laser (Nothing since 24 March 2019)
5. Andscacs (0.95 released 17 December 2018)
6. Fizbo (Nothing since 21 December 2017)

Division 2
1. Jonny (Version 8.1 used in TCEC for 6 consecutive seasons)
2. Chiron (Version 4 released on 10 January 2017. No version 5 yet).
3. Ginkgo (TCEC performance is worse over time. Not sure if it's still being developed)
4. ChessBrainVB (Nothing since 15 March 2019. Version 3.72 used in TCEC for 2 consecutive seasons)
5. Booot (Nothing since 8 July 2018)
6. Rofchade (All good for now).

Division 3
1. Fritz (Version 16 used in TCEC for 5 consecutive seasons)
2. Nirvana (Nothing since 11 August 2017)
3. Arasan (All good for now)
4. Texel (Seems stuck on version 1.08a13)
5. Vajolet2 (recently upgraded from 2.6 to 2.7)
6. Gull (dead since 2014)

The number of AB engines that are going to break out of Division 4 and get to Division 2 (meaning they place 1st or 2nd in Division 3) is zero. Neural Network engines on the other hand, using Lc0's progress, have no problem climbing to the Premier Division and staying there.

AllieStein began TCEC Season 15 (it's first season in TCEC) in Division 4, but ended up finishing 4th in the Premier Division ahead of Houdini 6.03.

If you dropped Antifish or Leelenstein into Division 4, they'd stroll into the Premier Division as well. It's seems that anything that is created using at least some part of Lc0, is automatically stronger than Fire 7.1 at the very least. For AB engine programmers, even ending the life of the engine at the same strength as Fire 7.1 is highly unlikely.

The barrier that no AB engine can get through is Fire level. Rofchade is the new kid on the block. Let's see how far it gets. Otherwise, TCEC is very stagnant unless NN engines (Lc0 clones) are your cup of tea.

[Ozymandias]

It's seems that anything that is created using at least some part of Lc0, is automatically stronger than Fire 7.1 at the very least. For AB engine programmers, even ending the life of the engine at the same strength as Fire 7.1 is highly unlikely.

Just allow the same latitude for SF derivatives and you'll have a plethora of engines capable of rivalling with the Lc0 derivatives.

The truth is that only two massive projects are getting past a common ceiling, reached by Houdini and Komodo almost two years ago (depending on the TC you'll find one or the other slightly ahead). Whether the MCTS version of the latter is able to get past that... only time will tell. Anything else getting there is highly unlikely.

[Gary Internet]

Just allow the same latitude for SF derivatives and you'll have a plethora of engines capable of rivalling with the Lc0 derivatives.

The truth is that only two massive projects are getting past a common ceiling, reached by Houdini and Komodo almost two years ago (depending on the TC you'll find one or the other slightly ahead). Whether the MCTS version of the latter is able to get past that... only time will tell. Anything else getting there is highly unlikely.

I agree with everything you've said. It's a shame because I like watching AB engines develop and improve over time, but it seems that after very quick early gains a point of diminishing returns is hit and people rightly don't want to devote countless weeks and months to tracking down a handful of extra Elo points.

[M ANSARI]

AB still have a place in chess ... they are the best calculators and usually will give you a sane opinion on a position. With NN engines it is still not there ... you really are never sure whether a +2.00 is an advantage or not. NN engines will need time to develop and in some positions they are nowhere near as strong as AB engine. I feel that for AB to continue, they need to take advantage of some of the developments of NN engines and maybe the strongest engine in a few years will be a hybrid engine using the best of AB and best of NN. This is not only due to development of software but also due to the development of hardware. At the moment AB engines use CPU cores and very little of the GPU ... and with NN it is the other way around. It makes sense from a hardware perspective to use both together ... and AI hardware is just getting started and will probably explode in strength in the very near future.

[Nay Lin Tun]

TCEC nightbot says!!

Capture.PNG

[mwyoung]

AB still have a place in chess ... they are the best calculators and usually will give you a sane opinion on a position. With NN engines it is still not there ... you really are never sure whether a +2.00 is an advantage or not. NN engines will need time to develop and in some positions they are nowhere near as strong as AB engine. I feel that for AB to continue, they need to take advantage of some of the developments of NN engines and maybe the strongest engine in a few years will be a hybrid engine using the best of AB and best of NN. This is not only due to development of software but also due to the development of hardware. At the moment AB engines use CPU cores and very little of the GPU ... and with NN it is the other way around. It makes sense from a hardware perspective to use both together ... and AI hardware is just getting started and will probably explode in strength in the very near future.

You are correct, but you must remember the "+2.00" is a human interpretation of Lc0 native evaluation. Lc0 does not score in fractions of pawns like AB engines. If the score is wrong or unreliable. It could be due more to the wrong formula being used to express Lc0 evaluation into a standard AB evaluation.

[Nay Lin Tun]

Credit Ali!!

Leela first encounter with Stockfish in last year!@

[Nay Lin Tun]

Leela Today!!

c3mLkQ9_d.jpg