Leela vs Stockfish nnue style

[h1a8]

Is there a significant difference in playing style between the two?
It's black Friday and I was going to either get a laptop with high end processor and low-mid gaming chip
Or one with medium-high processors with mid-high gaming chip. All the same price.
Or a desktop with max specs lol.

I know stockfish is stronger but Leela revolutionized chess with some radical new ideas in the opening and with strategy.

I know the nnue allows stockfish to operate similar to Leela. But how close? Do we need to invest in a high end graphics card to be able to get some good ideas for chess (through Leela)
Or can I just have a strong processor and use cfish (or a version that allows nnue only evaluation) to get similar results in ideas?

[Werewolf]

Because MCTS doesn't assume optimum play, Lc0 is a bit less dry.

[smatovic]

Can not tell about the neural network playing style, but SF uses AB search, more tactical, and Lc0 MCTS-PUCT/DAG search, more positional.

Some links:

LC0 vs. NNUE - some tech details...
https://www.talkchess.com/forum3/viewto ... =2&t=74607

Using LC0 with one or two GPUs - a guide
https://www.talkchess.com/forum3/viewto ... =2&t=70362

CPU Vector Unit, the new jam for NNs...
https://www.talkchess.com/forum3/viewto ... =2&t=75862

--
Srdja

[abgursu]

Is there a significant difference in playing style between the two?
It's black Friday and I was going to either get a laptop with high end processor and low-mid gaming chip
Or one with medium-high processors with mid-high gaming chip. All the same price.
Or a desktop with max specs lol.

I know stockfish is stronger but Leela revolutionized chess with some radical new ideas in the opening and with strategy.

I know the nnue allows stockfish to operate similar to Leela. But how close? Do we need to invest in a high end graphics card to be able to get some good ideas for chess (through Leela)
Or can I just have a strong processor and use cfish (or a version that allows nnue only evaluation) to get similar results in ideas?

You know who else revolutionalized chess? Stockfish.
Stockfish is nnue only evaluation. HCE got deleted for quite some time now. And unless you will only analyze positions without any tactical measurements, like fortressy closed positions, tactical strenght is important.
I don't understand people who can generalize it like lc0 is better positionally and stockfish is better tactically. Nope, for some positions Lc0 is better for sure, but for some positions Stockfish is better, "positionally". Same applies with tactical abilities too, sure often nn/mcts engines seem like they're worse on tactical strenght but sometimes they're better than stockfish at some kind of tactics (Remember the old Stoofvlees 1-0 Stockfish game from CCC 9 The gauntlet Quarterfinals round 432). Positional chess and tactical chess go hand in hand, you cannot seperate them completely with one cut. What results showing is that Stockfish plays better chess. Holds the hands better. Maybe you feel like you will use Leela to find positional novelties nobody has ever considered, well friend, you can do all that with Captain Stockfish and his clone army or other top level engines like Dragon, DragonMCTS(very good for new positional ideas!), Berserk, CSTal(also incredibly good with extraordinary ideas) etc. I'd specify my computer through my needs, not the engine I'll use mainly.

[h1a8]

Is there a significant difference in playing style between the two?
It's black Friday and I was going to either get a laptop with high end processor and low-mid gaming chip
Or one with medium-high processors with mid-high gaming chip. All the same price.
Or a desktop with max specs lol.

I know stockfish is stronger but Leela revolutionized chess with some radical new ideas in the opening and with strategy.

I know the nnue allows stockfish to operate similar to Leela. But how close? Do we need to invest in a high end graphics card to be able to get some good ideas for chess (through Leela)
Or can I just have a strong processor and use cfish (or a version that allows nnue only evaluation) to get similar results in ideas?

You know who else revolutionalized chess? Stockfish.
Stockfish is nnue only evaluation. HCE got deleted for quite some time now. And unless you will only analyze positions without any tactical measurements, like fortressy closed positions, tactical strenght is important.
I don't understand people who can generalize it like lc0 is better positionally and stockfish is better tactically. Nope, for some positions Lc0 is better for sure, but for some positions Stockfish is better, "positionally". Same applies with tactical abilities too, sure often nn/mcts engines seem like they're worse on tactical strenght but sometimes they're better than stockfish at some kind of tactics (Remember the old Stoofvlees 1-0 Stockfish game from CCC 9 The gauntlet Quarterfinals round 432). Positional chess and tactical chess go hand in hand, you cannot seperate them completely with one cut. What results showing is that Stockfish plays better chess. Holds the hands better. Maybe you feel like you will use Leela to find positional novelties nobody has ever considered, well friend, you can do all that with Captain Stockfish and his clone army or other top level engines like Dragon, DragonMCTS(very good for new positional ideas!), Berserk, CSTal(also incredibly good with extraordinary ideas) etc. I'd specify my computer through my needs, not the engine I'll use mainly.

Thank for your reply. I do not value 3000 elo tactical ability. That is something I will never be able to mimic or understand. I'll accept 1800-2200elo tactics with 3000+elo positional.

I didn't know how nnue works. I assumed it was a feature that mimics Leela (since Leela uses a neural network file for weights too).

In cfish there is an option to do hybrid (normal search with nnue evaluation), nnue only evaluation, and classical evaluation.
I assumed the nnue only evaluation will play similarly to Leela.

I'm only getting laptop for chess use (study, etc). No other needs.

Does Dragon Mtcs uses graphics chip like Leela for mtcs? Or does it uses only cpu?

[abgursu]

I don't know how strong or well informed a chess enthusiast you are, but I believe that you are a AlphaZero fan who read papers about it and try to simulate one. You did your research, found out about lc0 trying to simulate it, and then naturally want to use it. I'm coming this conclusion because all those things I said and the only thing interested you is Dragon using MCTS. You seem to informed only about mcts/nn and differences between cpu/gpu and still believe mcts is better than minimax or gpu engines are better than cpu or lc0 plays better "positional" chess like it's 2019.

Lc0 went on different roads from alphazero a long ago, like stockfish went different ways from it's past long ago. There are no clear superiority between them in equal normal opening/midgame positions, but we can say Stockfish is superior a little in biased openings & following positions and endgames due to it's search. Lc0 is a little superior in some themes like fortressy positions and trapped pieces. Not positional/tactical nor their elos, not that simple. Some themes. Overall, these differences make Stockfish a little better generally, noone is trying to mimic anyone. Both are so strong on a good hardware that you cannot compare them very good. And... both are stronger than the 2017-18 AlphaZero.

Coming to your question, DragonMCTS uses CPU not GPU. Good alternative engine for discovering different ideas.

If you want my suggestion on the topic, I say go fish or go home. The only top level engine you can use which uses gpu is Leela, while every other one is only affected by how strong your cpu's are. Buying a whole computer only to use leela is... pointless.

[syzygy]

I don't understand people who can generalize it like lc0 is better positionally and stockfish is better tactically. Nope, for some positions Lc0 is better for sure, but for some positions Stockfish is better, "positionally".

These statements follow naturally from the differences between the engines. Since Lc0 searches far fewer nodes per second, it necessarily relies more on its heavier static evaluation than SF. If you have to rely more on your static evaluation, you select moves more on "positional" than on "tactical" grounds.

I do agree that this "positional" vs"tactical" distinction does not necessarily translate into Lc0 playing like Karpov and SF playing like Kasparov. But in the same highly tactical position, Lc0 will be more likely to miss the best move.

[syzygy]

Buying a whole computer only to use leela is... pointless.

If someone is primarily interested in AlphaZero-like engines, then Lc0 is the way to go.
The computer can then also be used for other machine-learning type applications. Or for playing videogames.

[h1a8]

I don't know how strong or well informed a chess enthusiast you are, but I believe that you are a AlphaZero fan who read papers about it and try to simulate one. You did your research, found out about lc0 trying to simulate it, and then naturally want to use it. I'm coming this conclusion because all those things I said and the only thing interested you is Dragon using MCTS. You seem to informed only about mcts/nn and differences between cpu/gpu and still believe mcts is better than minimax or gpu engines are better than cpu or lc0 plays better "positional" chess like it's 2019.

Lc0 went on different roads from alphazero a long ago, like stockfish went different ways from it's past long ago. There are no clear superiority between them in equal normal opening/midgame positions, but we can say Stockfish is superior a little in biased openings & following positions and endgames due to it's search. Lc0 is a little superior in some themes like fortressy positions and trapped pieces. Not positional/tactical nor their elos, not that simple. Some themes. Overall, these differences make Stockfish a little better generally, noone is trying to mimic anyone. Both are so strong on a good hardware that you cannot compare them very good. And... both are stronger than the 2017-18 AlphaZero.

Coming to your question, DragonMCTS uses CPU not GPU. Good alternative engine for discovering different ideas.

If you want my suggestion on the topic, I say go fish or go home. The only top level engine you can use which uses gpu is Leela, while every other one is only affected by how strong your cpu's are. Buying a whole computer only to use leela is... pointless.

I can't understand Stockfish moves most of the time. There are heavily based in tactics (far above human level). my tactical ability will never be above 2200level so most of the moves are over my head. leela (or mcts engines i assume), on the other hand, moves are primarily based on static evaluation. this is easy for me to learn from and get better at chess. the goal isn't to have the strongest engine but the engine EASIEST to learn from (where tactical ability is at a minimum). Ill try to get the laptop that has better cpu (weaker gpu) and try dragon mcts. that will save me some money. just hoping the moves are not heavily based on deep tactical search.

[Uri Blass]

I don't know how strong or well informed a chess enthusiast you are, but I believe that you are a AlphaZero fan who read papers about it and try to simulate one. You did your research, found out about lc0 trying to simulate it, and then naturally want to use it. I'm coming this conclusion because all those things I said and the only thing interested you is Dragon using MCTS. You seem to informed only about mcts/nn and differences between cpu/gpu and still believe mcts is better than minimax or gpu engines are better than cpu or lc0 plays better "positional" chess like it's 2019.

Lc0 went on different roads from alphazero a long ago, like stockfish went different ways from it's past long ago. There are no clear superiority between them in equal normal opening/midgame positions, but we can say Stockfish is superior a little in biased openings & following positions and endgames due to it's search. Lc0 is a little superior in some themes like fortressy positions and trapped pieces. Not positional/tactical nor their elos, not that simple. Some themes. Overall, these differences make Stockfish a little better generally, noone is trying to mimic anyone. Both are so strong on a good hardware that you cannot compare them very good. And... both are stronger than the 2017-18 AlphaZero.

Coming to your question, DragonMCTS uses CPU not GPU. Good alternative engine for discovering different ideas.

If you want my suggestion on the topic, I say go fish or go home. The only top level engine you can use which uses gpu is Leela, while every other one is only affected by how strong your cpu's are. Buying a whole computer only to use leela is... pointless.

I can't understand Stockfish moves most of the time. There are heavily based in tactics (far above human level). my tactical ability will never be above 2200level so most of the moves are over my head. leela (or mcts engines i assume), on the other hand, moves are primarily based on static evaluation. this is easy for me to learn from and get better at chess. the goal isn't to have the strongest engine but the engine EASIEST to learn from (where tactical ability is at a minimum). Ill try to get the laptop that has better cpu (weaker gpu) and try dragon mcts. that will save me some money. just hoping the moves are not heavily based on deep tactical search.

Watching stockfish-lc0 games in tced I see that
Stockfish's moves and lc0's moves are often the same moves and often the pv is the same for a lot of plies so I see no importance for the question if they are based on tactical calculations or positional evaluation.

If you are interested only in lc0's move at low number of nodes per move so it has not a good search then
Lc0 has no simple evaluation so I see no reason what is the reason to think that you can learn from the evaluation and improve in chess.