What Is The Horizon Effect?

Sort:
Avatar of superchessmachine

I have heard the term horizon effect when reading about computer's weaknesses. What is the horizon effect?

Avatar of IMKeto

The Horizon Effect is caused by the depth limitation of the search algorithm, and became manifest when some negative event is inevitable but postponable. Because only a partial game tree has been analyzed, it will appear to the system that the event can be avoided when in fact this is not the case. Beside obligatory quiescence searchextensions, specially check extensions are designed to reduce horizon effects.

Avatar of superchessmachine

Cant you just give the computer more hash and time to calculate?

Avatar of KZNinjago
What is the “ HORIZON AFFECT ? “
Avatar of superchessmachine

Thats what I want to know

Avatar of testaaaaa

old pcs had problems as black in the stonewall attack because they did not see the mate coming 

Avatar of superchessmachine

You know that because you read the article but yes you are right.

Avatar of SmithyQ

Example: Suppose there's a slow but unstoppable attack coming.  It's not forcing, it takes 12 moves but Black cannot stop it via any move order.  The computer only realizes the attack when it reaches a depth of ply 20 (say).

Once the computer hits ply 20, it realizes it is busted.  It then tries to find an alternative.  If it sacs a pawn, makes a spite check or does something that causes White to delay his attack (inserting a5 and a4, for instance), this pushes the realization of a buster position from ply 20 to ply 24 (say).  The computer evaluates this as fine, because the disaster has been pushed 'over the horizon'.  If the computer gets to ply 24, it will again see the attack is winning and again try to delay it by pushing it further away.

This can also occur with positional sacrifices (not realizing the long-term compensation) and fortresses (not realizing the defence cannot be broken).  Traditional engines can keep getting faster and push the horizon further away (ply 50, say), but it won't get rid of it.  The new breed of AI engines similar to AlphaZero may change this, but we'll see.

Avatar of yureesystem
IMBacon wrote:

The Horizon Effect is caused by the depth limitation of the search algorithm, and became manifest when some negative event is inevitable but postponable. Because only a partial game tree has been analyzed, it will appear to the system that the event can be avoided when in fact this is not the case. Beside obligatory quiescence searchextensions, specially check extensions are designed to reduce horizon effects.

 

 

 

Excellent explanation!!

 

Avatar of IMKeto
superchessmachine wrote:

Thats what I want to know

Think of it this way.  With black holes, you have a horizon effect.  You cant see anythig once it enters the black hole, but you can see everything right up to the moment it enters the black hole.

Avatar of superchessmachine
Carpe_noctum wrote:
IMBacon wrote:

The Horizon Effect is caused by the depth limitation of the search algorithm, and became manifest when some negative event is inevitable but postponable. Because only a partial game tree has been analyzed, it will appear to the system that the event can be avoided when in fact this is not the case. Beside obligatory quiescence searchextensions, specially check extensions are designed to reduce horizon effects.

after algorithm I stopped reading

Lol. I read the whole thing

Avatar of superchessmachine
SmithyQ wrote:

Example: Suppose there's a slow but unstoppable attack coming.  It's not forcing, it takes 12 moves but Black cannot stop it via any move order.  The computer only realizes the attack when it reaches a depth of ply 20 (say).

Once the computer hits ply 20, it realizes it is busted.  It then tries to find an alternative.  If it sacs a pawn, makes a spite check or does something that causes White to delay his attack (inserting a5 and a4, for instance), this pushes the realization of a buster position from ply 20 to ply 24 (say).  The computer evaluates this as fine, because the disaster has been pushed 'over the horizon'.  If the computer gets to ply 24, it will again see the attack is winning and again try to delay it by pushing it further away.

This can also occur with positional sacrifices (not realizing the long-term compensation) and fortresses (not realizing the defence cannot be broken).  Traditional engines can keep getting faster and push the horizon further away (ply 50, say), but it won't get rid of it.  The new breed of AI engines similar to AlphaZero may change this, but we'll see.

Thanks for explaining! You and @IMBacon had good explanations

Avatar of superchessmachine
IMBacon wrote:
superchessmachine wrote:

Thats what I want to know

Think of it this way.  With black holes, you have a horizon effect.  You cant see anythig once it enters the black hole, but you can see everything right up to the moment it enters the black hole.

So like:

By the time the computer realizes that it is ded it is too late?

Avatar of superchessmachine

oh! I see

Avatar of Preggo_Basashi
superchessmachine wrote:

What Is The Horizon Effect?

The engine has to stop calculation at some point.

Unluckily for it, sometimes it's just before a key move (like checkmate!)

But since it can calculate really deep, it tends to correct itself before the position is beyond saving. Years ago when engines weren't as good, you could build up a positional attack and they wouldn't realize until it was too late.

 

 

superchessmachine wrote:

Cant you just give the computer more hash and time to calculate?

It has to stop somewhere.

It can't calculate until mate in 100% of lines unless you give it access to memory whose hardware would literally be planetary sized (if you store 1 position on the size of 1 atom, then how many atoms, you count them up, and boom, you'd need hardware larger than earth itself lol).

Avatar of superchessmachine
Preggo_Basashi wrote:
superchessmachine wrote:

What Is The Horizon Effect?

The engine has to stop calculation at some point.

Unluckily for it, sometimes it's just before a key move (like checkmate!)

But since it can calculate really deep, it tends to correct itself before the position is beyond saving. Years ago when engines weren't as good, you could build up a positional attack and they wouldn't realize until it was too late.

 

 

superchessmachine wrote:

Cant you just give the computer more hash and time to calculate?

It has to stop somewhere.

It can't calculate until mate in 100% of lines unless you give it access to memory whose hardware would literally be planetary sized (if you store 1 position on the size of 1 atom, then how many atoms, you count them up, and boom, you'd need hardware larger than earth itself lol).

[Amazon Cloud Computing Intensifies]

Avatar of superchessmachine
Preggo_Basashi wrote:
superchessmachine wrote:

What Is The Horizon Effect?

The engine has to stop calculation at some point.

Unluckily for it, sometimes it's just before a key move (like checkmate!)

But since it can calculate really deep, it tends to correct itself before the position is beyond saving. Years ago when engines weren't as good, you could build up a positional attack and they wouldn't realize until it was too late.

 

 

superchessmachine wrote:

Cant you just give the computer more hash and time to calculate?

It has to stop somewhere.

It can't calculate until mate in 100% of lines unless you give it access to memory whose hardware would literally be planetary sized (if you store 1 position on the size of 1 atom, then how many atoms, you count them up, and boom, you'd need hardware larger than earth itself lol).

I see. That part was clear before.

 

Now I understand the whole concept. But wouldn't the computer out calculate a human? I suppose this only applies for positional weakness's and attack's yeah?

Avatar of Preggo_Basashi
superchessmachine wrote:
Preggo_Basashi wrote:
superchessmachine wrote:

What Is The Horizon Effect?

The engine has to stop calculation at some point.

Unluckily for it, sometimes it's just before a key move (like checkmate!)

But since it can calculate really deep, it tends to correct itself before the position is beyond saving. Years ago when engines weren't as good, you could build up a positional attack and they wouldn't realize until it was too late.

 

 

superchessmachine wrote:

Cant you just give the computer more hash and time to calculate?

It has to stop somewhere.

It can't calculate until mate in 100% of lines unless you give it access to memory whose hardware would literally be planetary sized (if you store 1 position on the size of 1 atom, then how many atoms, you count them up, and boom, you'd need hardware larger than earth itself lol).

I see. That part was clear before.

 

Now I understand the whole concept. But wouldn't the computer out calculate a human? I suppose this only applies for positional weakness's and attack's yeah?

Yeah, the engine is only weak when it comes to long term things, like endgames, or even attacks on the king, if they're positionally based.

Engines play 100% perfectly in the short term. You know, like mate in 2 or 3 it will always find it. But the further away they have to calculate, the less accurate they are. Meanwhile humans can rely on things we've learned in the past, and we can imagine the future without calculating every move in between to get there.

 

A simple example of imagination an engine can't do

 

Now a human knows, that due to the pawn structure (above) that if you remove all the pieces except the pawns and kings the game is a win for white. It's a win because white's kingside can make a passed pawn, and black's queenside majority can't (due to the doubled pawns).

 

So as humans, we can judge certain piece trades (every trade brings us closer to the lost/won endgame). We can do this starting even on move 5, and imagine what sorts of ramifications they might have on move 50, or move 100, etc. This is not possible for an engine.

 

This doesn't have a big effect on the immediate evaluation, but it's just an example of how humans can easily skip all the moves in between, and imagine something far in the future right away.

Avatar of KingSullian

Preggo_Basashi wrote:
 

It has to stop somewhere.

It can't calculate until mate in 100% of lines unless you give it access to memory whose hardware would literally be planetary sized (if you store 1 position on the size of 1 atom, then how many atoms, you count them up, and boom, you'd need hardware larger than earth itself lol).

[Amazon Cloud Computing Intensifies]

 

I would place money that you have read or heard of a solution, a proposition, a theoretical theory of sorts in physics that does exactly that. If you would fill up the entire universe with information such as 1,0"s/high/lows, off/ons..etc. Just this data would slam the known visible universe,

the 4 dimensions into an equilibrium, a parity, A thermodynamic equality of nucleosynthesis.. BANG!    

Avatar of RonaldJosephCote

Sorry for the 1 post hi-jack......look at it as an example of the horizon effect.wink.png                                                           null