Tactics cannot be very important for chess

[dragontamer5788]

Since LC0 performed so terribly (I gave 12 minutes per position on a test set meant to run for 15 seconds per position) and is beaten by all the strong tactical engines by a landslide, I can only conclude that tactics are not very important for the game of chess.

LC0 has only a few-hundred thousand nodes per second. This means it is significantly stronger at tactics than any human player, but still far slower than the millions or even hundred-millions nodes-per-second of classic A/B engines.

Tactics remain important: but LC0 has "enough" tactics to play decently. Where it wins is that LC0 has very strong positional play / long-term skill in chess.

I suspect that I made a bad net choice, so I am replaying.
But I gave LC0 48 times more time per position, and the results were still terrible.
So I suspect your diagnosis is correct.
I also expect the bigger net to do better

Smaller nets probably will be better in tactics, while bigger nets will probably be stronger for positional play. A bigger net means spending more time evaluating the neural net... which necessarily means having fewer-and-fewer nodes-per-second.


Case in point: a small net may have 10-microseconds per node (100,000 nodes per second), while a big net may have 50 microseconds per node (20,000 nodes per second).

Not necessarily, there are policy and value parts affecting the tactics, not just NPS. It depends on say time control too. I have mixed results on that on tactical suites like Arasan and trimmed WAC. Bigger nets seem to improve on tactics faster with longer TC than smaller nets. But many smaller nets start from better tactics at very short TC.

Yeah, I forgot about the policy part when I wrote that post earlier.

One more potential issue: GPUs need a very large number of threads before they're practical. The NVidia 1080 Ti has 3584 CUDA-cores, so you need at least 3584 SIMD-threads before you really utilize the system (1-thread will run just as quickly as 3584-threads). Furthermore, GPUs don't have branch-prediction, memory-prefetching, or other high-performance features that CPU-programmers use.

Instead: GPUs use a hyperthread-like feature where they can quickly switch between many threads while waiting for memory requests. For NVidia, IIRC there can be up to 32x warps per SM, or roughly 8x logical-threads per physical CUDA-thread. Max occupancy would therefore be 28672 SIMD-threads per NVidia 1080 Ti.

I figure the neural nets are evaluated in parallel to some degree with MCTS-virtual loss + asynchronous compute. But a larger neural net will force a larger amount of work, which would make the GPU more efficient. Having lots-and-lots of work helps a lot. In this case, its not that the "bigger net" will run faster, its that you have so many CUDA-cores that the "bigger net" might not really lose much in terms of nodes-per-second compared to a smaller net.

[Uri Blass]

Many, many chess games are won with a gradually increasing eval for the winning side. So there really is no tactical moment where the game is won, there is just increasing pressure and eventually conversion into a mate. Carlsen is this type of player, generally: he is just really good at squeezing a small advantage into a win.

Once in while you do need tactical acuteness, because your opponent has a deep threat you must counter, or there's a winning shot that you really need to find. But you can go a lot of games without actually needing this.

By defintion there is always a tactical moment(I assume for the discussion that opening position is a draw by perfect play)
Maybe computers are too weak to see it today but there is a move that changed the theoretical result from draw to loss so we can say it is a tactical mistake.

[towforce]

I did an experiment (perhaps flawed because I used this net:128x10-2019_1026_0142_45_230.pb) and LC0 is really terrible at tactics on a Nvidia TI-1080.
The engine crashed once when I was analyzing, so I restarted at the point where analysis left off, and it did not crash a second time.

I am re-running the test with a 256 net 256x20-2019_0726_0906_48_558.pb

Since LC0 performed so terribly (I gave 12 minutes per position on a test set meant to run for 15 seconds per position) and is beaten by all the strong tactical engines by a landslide, I can only conclude that tactics are not very important for the game of chess.

On the other hand, if they are important, and they can somehow be compiled into an LC0 net, perhaps LC0 would become utterly unbeatable

It looks to me as though we are entering a third historic phase of computer chess: up until 1997, it was common to hear people say things like, "They won't get GM level in my lifetime". Since then, ever increasing computer power has rewarded quick evaluations and deep search, It was common to find programmers throwing knowledge out of the eval, as depth of search often rendered it irrelevant.

Suddenly, over 20 years on from Deeper Blue v Kasparov, we're finding that beyond a certain level, positional play (and hence knowledge) is a key driver - and we have a technology to deliver that knowledge: big data.

For a game played on a small board with a small number of pieces, it turns out that chess has been unbelievably good at guarding its deepest secrets against an absolutely stunning advancement in computer technology since WWII. Given that a mathematical proof that chess is a drawn game looks out of reach right now, maybe a combination of knowledge and search can take us to the highest point which can be reached: a player that never gets beaten.

[Laskos]

Many, many chess games are won with a gradually increasing eval for the winning side. So there really is no tactical moment where the game is won, there is just increasing pressure and eventually conversion into a mate. Carlsen is this type of player, generally: he is just really good at squeezing a small advantage into a win.

Once in while you do need tactical acuteness, because your opponent has a deep threat you must counter, or there's a winning shot that you really need to find. But you can go a lot of games without actually needing this.

By defintion there is always a tactical moment(I assume for the discussion that opening position is a draw by perfect play)
Maybe computers are too weak to see it today but there is a move that changed the theoretical result from draw to loss so we can say it is a tactical mistake.

Can we actually operate with vaguer notions? Say, by taste and empirically, for Stockfish, tactics is usually less than 40 plies deep shot to pure eval level, strategy is usually above 60 plies deep shot to pure eval level? You often try to clarify to mathematical and logical level things in this dirty world of very imperfect chess engines, having so many imperfections and unknowns that the study of them is often more akin to anatomy than to math.

[Dann Corbit]

Given the success of LC0, I suggest that strategic mistakes are far more probable than tactical mistakes.
LC0 is positively fragile, tactically. And yet it is easily as strong as Stockfish.
This is not superb evidence and largely conjecture, but I suppose that most mistakes are actually strategic mistakes.

[Laskos]

Given the success of LC0, I suggest that strategic mistakes are far more probable than tactical mistakes.
LC0 is positively fragile, tactically. And yet it is easily as strong as Stockfish.
This is not superb evidence and largely conjecture, but I suppose that most mistakes are actually strategic mistakes.

I tend to agree with that. Never mind Stockfish absolute tactical superiority and bad results against Lc0 in normal games on my PC. Even when comparing different Lc0 nets, the results on my "positional test suites" are much more relevant for determining the strength than those on "tactical test suites". Do you remember that LeelaFish Lc0 + SF hibrid? It was vastly superior tactically compared to regular Lc0, and only a little inferior positionally. All in all, it gave no much improvement over Lc0. Decreasing the positional prowess of Lc0 seems a bad idea strength-wise.

My impression for years now is that stronger the engines are at tactics, less important it becomes. Important tactical shots seem to cluster at shallow level (say depth), and then decrease in importance. And what tactics Lc0 knows seems quite enough to avoid such shallow mistakes, not always, but often enough to beat Stockfish on my PC in matches consisting of normal games (say, not picked very tactical openings).

[duncan]

Given that a mathematical proof that chess is a drawn game looks out of reach right now, maybe a combination of knowledge and search can take us to the highest point which can be reached: a player that never gets beaten.

So that would be a player that can never be beaten by itself even if given 10 times more time. Would not rule it out within 10 years.

[Uri Blass]

Many, many chess games are won with a gradually increasing eval for the winning side. So there really is no tactical moment where the game is won, there is just increasing pressure and eventually conversion into a mate. Carlsen is this type of player, generally: he is just really good at squeezing a small advantage into a win.

Once in while you do need tactical acuteness, because your opponent has a deep threat you must counter, or there's a winning shot that you really need to find. But you can go a lot of games without actually needing this.

By defintion there is always a tactical moment(I assume for the discussion that opening position is a draw by perfect play)
Maybe computers are too weak to see it today but there is a move that changed the theoretical result from draw to loss so we can say it is a tactical mistake.

Can we actually operate with vaguer notions? Say, by taste and empirically, for Stockfish, tactics is usually less than 40 plies deep shot to pure eval level, strategy is usually above 60 plies deep shot to pure eval level? You often try to clarify to mathematical and logical level things in this dirty world of very imperfect chess engines, having so many imperfections and unknowns that the study of them is often more akin to anatomy than to math.

The question is what is the definition of tactics that you choose.
I think that it is wrong to define tactics based on stockfish's view.

I thought that you can define tactics as a combination that win material or mate in at most 40 plies and even here it is not clear because maybe you can force winning material but the opponent has some compensation so practically you win nothing.

Stockfish also has evaluation function and can evaluate positional factors as more important than material.

Maybe you can define tactics as a move that humans can be convinced by analysis with engines that it force a win(or a draw) when it is clear that a different move is worse.
Again it is not a clear definition

[duncan]

Maybe you can define tactics as a move that humans can be convinced by analysis with engines that it force a win(or a draw) when it is clear that a different move is worse.
Again it is not a clear definition

Could you not also define strategy that way. ?Why not a tactical move is a move that within a limited amount of moves gains material without any equalising material or positional compensation by the opponent. (Avoiding strategy is long distance tactics issue.)

[Dann Corbit]

I recall a time when a chess engine would think all night to achieve 17 plies.
How would we have defined tactical and strategic then?

I think the definition is a moving target.

But I would put it this way:
A strategic move is made using first principles without seeing a definite win of material or game outcome.
A tactical move is made for a known purpose, either a win in material or mate or draw achieved.
Hence, for instance, giving up a full piece might be either tactical or strategic.