Alpha-Emanuel-Beta Search -- Enormous Improvement of Alpha-Beta!

[R. Tomasi]

Okay, I can see how a form of this idea can work. It is white to move and we are about to call for a 40 ply deep search. We also have a best move from ply depth 39. We like that move because it is good enough. Therefore we only need to know that there is no deep tactic that ruins that move. So we only need to search that one move to depth 40. Then we need to search all black moves to try to find a refutation. Next we only need to search one white move. Etc. How we get that one white move to try can be from the hashtable or just a guess. Only if the one white move fails does it revert back to being pure AB. What if at any depth remaining there is no obviously best move? Then do a more shallow search to find one.

The whole idea boils down to good enough for one side. Now we have to define good enough. Good enough has to be better than two moves earlier or we can end up not making progress.

Logically AB is looking for the truth of the position and the absolute best move. Therefore Emanuel not looking for absolute truth could reduce the search tree immensely just by looking for good enough. The ELO gain if there is one would come from searching a good enough move far deeper to make sure it really is good enough!

Trouble I have with that idea is twofold:

• From which depth distance to the root node onwards will I switch to not looking at my other options anymore? Clearly, if I use the algorithm right from the root, I will essentially only look at one move, and I am prepared to bet it's a terrible one...
• If I stop looking for optimal and satisfy myself with "good enough", then my evaluation function, which is supposed to guide me towards a stronger position will not be able to do that job.

[JohnWoe]

I got +200 ELO !!!
Many thanks!

I assume that you understand the algorithm and are not making a joke. But, then why is your comment being ignored? If it is just a joke then why haven't you got a "nice one" or "lol". And if you are not joking ... ? I feel like a fish watching a shiny object skimming across the surface of the water. Can you please explain?

Of course joking.
The "algorithm" only assumes 1 black move when winning or smt.
I didn't try the "algorithm" on my program.
Some new non-recursive algorithm would be nice. Not some +0.5 ELO reduction idea ...

[Chessnut1071]

and player A still comes out a winner, then we have proven that A will win.

"No sh*t, Sherlock!"(tm)

Ok.. so you understand my algorithm? Can you tell the others?

Obviously you haven't tried that. If you prune, I can always find a mating sequence that will get past your algorithm, even if it's low probability. Trust me, I use 2,560 mating puzzles and there is very few systems I haven't tried yet. Alpha-beta is the only one that correctly solves the least move checkmate on every one.

Oh ok, no as I wrote above there is no guarantee of proving the bound in the minimum number of moves. But that's kind of a niche use case. More often you want to know one player is guaranteed a mate or certain piece advantage, not whether it happens in 5 or 7 moves.

[Chessnut1071]

and player A still comes out a winner, then we have proven that A will win.

"No sh*t, Sherlock!"(tm)

Ok.. so you understand my algorithm? Can you tell the others?

Obviously you haven't tried that. If you prune, I can always find a mating sequence that will get past your algorithm, even if it's low probability. Trust me, I use 2,560 mating puzzles and there is very few systems I haven't tried yet. Alpha-beta is the only one that correctly solves the least move checkmate on every one.

Oh ok, no as I wrote above there is no guarantee of proving the bound in the minimum number of moves. But that's kind of a niche use case. More often you want to know one player is guaranteed a mate or certain piece advantage, not whether it happens in 5 or 7 moves.

Obviously, we're working on different objectives. I think we can both agree that alpha-beta without pruning is optimal. It all comes down to hash tables and evaluation functions. I could care less about ELO scores, since the future is quantum computing. Future speeds will be magnitudes faster and optimal engines will eventually eclipse most of the stuff in vogue today.

[klx]

Okay, I can see how a form of this idea can work. ... Emanuel not looking for absolute truth could reduce the search tree immensely just by looking for good enough.

That's exactly right. This guy gets it!

[klx]

Ok.. so you understand my algorithm? Can you tell the others?

You have just re-invented the Shannon B type strategy from 1949. Alpha-Beta engines do implement a mix of Shannon A (brute force) and Shannon B (selective search), it's called alpha-beta with extensions and unsafe pruning, plus quiescence which is also selective.

Never heard of these concepts, but just read up on them, and no that's not what I'm doing. Shannon type strategy still searches the same for both players. Alpha-Emanuel-Beta, in contrast, forms a hypothesis before the search which player will win or prove the bound for, and only uses the selectivity for that player.

I suppose it would be something like Shannon B type strategy for one player and A type for the other. Shannon never thought of that. I did.

[klx]

Just out of curiosity are you of portuguese or spanish language origins?

Bariloche, Argentina. So "Spanish" if you will, but I trace my ancestry to the northern parts of Italy (Lombardia), with a subtle hint of German.

[klx]

if I use the algorithm right from the root, I will essentially only look at one move, and I am prepared to bet it's a terrible one...

Well you would not use it from the root. For the first couple of plies it is of no use since the search is so fast anyway, and in fact it is likely to guide the search wrong. In my experiments, I have invoked this pruning only after several plies of iterative deepening.

Let me remind everyone that this is not just some wild theory. I have written a proof of concept of this, and seen super massive gains.

[Daniel Shawul]

Ok.. so you understand my algorithm? Can you tell the others?

You have just re-invented the Shannon B type strategy from 1949. Alpha-Beta engines do implement a mix of Shannon A (brute force) and Shannon B (selective search), it's called alpha-beta with extensions and unsafe pruning, plus quiescence which is also selective.

Never heard of these concepts, but just read up on them, and no that's not what I'm doing. Shannon type strategy still searches the same for both players. Alpha-Emanuel-Beta, in contrast, forms a hypothesis before the search which player will win or prove the bound for, and only uses the selectivity for that player.

I suppose it would be something like Shannon B type strategy for one player and A type for the other. Shannon never thought of that. I did.

Weren't you harping about how it is a safe/sound pruning strategy that improves alpha-beta enormously ??
Now you say it is shannon type B for one side, when I already pointed to you there is nothing better than alpha-beta that is still safe.

Question to others: Are you so desperate for "techinical discussions" that you have to work to come up with something to engage with him ?
The original propostiion was, "a safe/sound pruning strategy that gives exponential reduction over alpha-beta".
Instead of calling this nonsense out, you helped him divert to a different claim .. what a shit show.

[smatovic]

@Daniel: this guy is talking obv. nonsense since his last three Emanuel-inventions posted here. And he doesn't get it. So let's have som fun and figure out together how we get this Emanuel Pruning thingy running?

--
Srdja