Knights or Bishops? What AlphaZero Thinks

No, this isn't a rehash of a tired old debate.

The conventional wisdom is that bishops are better than knights and whenever someone posts asking this question, the vast majority of responses claim that bishops are better, with some even adding some nuanced caveats.

Well, if anyone cares, AlphaZero has spoken. As the chart below shows, the AI values both essentially identically. With less training, A0 slightly prefers bishops, but the more it trains, the more it comes to value them equally. I've always argued that at lower ratings, knights are more dangerous because knight geometry is hard to get a good handle on. Even though at higher ratings, masters strongly prefer bishops to knights, maybe it's because they don't understand knight geometry as well as AlphaZero does. Of course none of this really matters for mere mortals, I just thought it was interesting.

I think the current meta is that while a knight is equal to a bishop, there is an advantage in having 2 bishops vs 2 knights or bishop and knight, by about half a pawn.

The idea is that the two bishops complement each other. 2 knights have some redundancy. So for example there could be one good square that both knights want to occupy, but this is never the case with 2 bishops.

I have adjusted my playing to reflect this. I will fight for the two bishops, but if that is no longer possible I will treat knight and bishop as equals.

This assumes of course there is nothing about the position to favor one over the other. For example, what color are your pawns compared to the bishop, etc. This can also lead you to the knight over the bishop or vice versa.

this of course assumes you are at least at a skill level where you don’t have trouble spotting knight moves.

I teach beginners kids chess, and I tell them bishop and knight are worth 3 pawns.

In my “advanced” class (nothing anyone here would consider advanced!) I will point out good bishops and bad bishops, and open vs closed positions for knights and bishops, but I won’t talk about things like “half a pawn”. I just say “is a little better or worse”

I have the book “game changer”, I have only gone through the first chapter. I like how alpha zero plays.

Intersection of programming and chess interesting to me because career and hobby in one.

AFAIK (book hasn’t discussed this yet) Alpha zero does not actually carry a specific value of the pieces, it is all situational.

Having a static piece value is a crutch for us relatively simple minded primates.

Well, the chart assumes an overall preference in all situations, of course. One would think that if A0 even showed a slight preference for the bishop pair, it would show up in the overall preference since it’s a subset. Think about it, if A0 showed a preference for the pair, but the overall preference is even, then a slight preference for knights in other situations would be needed to create an overall preference balance.

My guess, and this is just a guess, is that A0 can look more deeply into a position than any human can and can see where a knight can be in, say, six moves far better than any human can. Knights are not blocked by pawns, so they have a lot more possible squares they can reach than bishops.

This is all just reckless speculation, though. I do think this just ends the argument that one is inherently better than the other, and that it always depends on context and the strength of the player.

It’s also interesting that A0 values rooks slightly more than 5 and queens closer to 10 than 9.

I think Alpha Zero is full of ****. 

Remember Alpha Zero gets no help from humans, so while a conventional chess engine (which completely destroy Alpha Zero every time) accepts tweaking to treat bishops as better, Alpha Zero does not have this feature. 

On another note I think in blitz knights are often practically-speaking equal or even better - even at higher levels. Because at the lower time limits people can't examine all the knight moves and they can be very dangerous. 

@uhohspaggettio1

I think you miss the point. In conventional engines, humans decide a priori what the value of each piece is. A0 plays against itself, decides what works and doesn't work, then shows a preference based on machine learning. No humans decide in advance for it, therefore it doesn't suffer from human biases.

Clearly self-learning AI has proven to be a superior way to create the strongest engines. If you believe that A0 is full of poo, then you're very clearly in the minority.

I actually can’t read the chart on my phone. Are you saying A0 values knights and bishops exactly even? It is there a discrepancy that at least in theory could come about as a preference for a bishop pair that humans have?

If this is some artifact of how engines play then it still in practice may not be that applicable to humans. GM master practice has a preference for the bishop pair. This is well over a century of human chess theory, and being human myself, I would tend to stick with that. Engines do play quite differently, in the same way that beginners and masters play differently, what applies to one may not apply to the other.

I do find A0 valuations fascinating, but I don’t necessarily think humans should accept it as guides for our play.

I think I read somewhere A0 values rooks in general as being 5.5 pawns, a little higher than human chess theory says.

 

I'm not missing any point. I think you don't understand my post, I made it as clear as I could and I think you might understand if you go back and read it again.  

Contrary to your statements self-learning AI by itself stands no chance against computer engines that have had human intervention in their programming, "human bias" as you call it farcically. If you doubt this then you are really ignorant about computer chess and shouldn't be making any comments on it. 

 

I believe that a piece's value, regardless of its initial value, is determined by its position.

https://www.chessgames.com/perl/chessgame?gid=1012123

In this game, blacks white squared bishop is worth 0.

It all depends on the games position. 

Yes of course piece valuation is situational, but since humans aren’t infinite calculating engines capable of precisely adjusting valuations on the fly we have the crutch of general statistical long term approximate values to help guide us, which stockfish I think also uses. I am pretty sure A0 does not.

I think purely situational values would be the correct way to think if we were infinitely intelligent. But since most humans (cough) aren’t, we need that helper value.

(In case it isn’t obvious, when I say “situational value”, that phrase is carrying a lot of weight)

I don't think it takes infinite calculations to predict situational value unless you need to find it exactly and accurately. It needs some positional understanding and a bit of a common sense. At the end of the day, the value is not the important matter here. Important is whether we are able to calculate better in a position and make decision quicker. It is just for the sake of aiding in calculation and I think we already have a grasp of it by now. 

 

I mean what's the point of knowing bishop is slightly better than Knight if the value changes position wise.

The point is that we humans are not that great at calculating precisely what the values are in a position.

If we are in a generic early middle game game and I can trade my bishop for your rook, I am going to do it provided I can’t see an immediate tactical problem.

I am doing that because I know in the long run your rook is worth 5 points and my bishop 3, because I read it somewhere.

If I didn’t read it somewhere, and so didn’t know their general value, would I trade it based on what I can calculate their precise value now and in the future to be? Oh hell no lol. My valuation power is nowhere close to doing that.

I get what you mean. Like I said, I don't think it’s possible to find values of each position and remember them. Especially, not accurate value. All we do is deduce from our positional understanding. So I guess we are on the same page here. 

However, in terms of Knight and bishop, they are so close to each other in terms of value that there is no point in fixing them as slightly better. In a closed position, I'd try to trade off a bishop for a Knight (depending on the position). For an example : in French defense, Black would be glad to trade off their bad bishop. 

 

So I know you are saying we need a fixed value so that we can have a grasp of the calculation and I am saying that yes we need it but we cannot claim it objectively that rook has worth 5 values all the time. All we can say is that it’s just a "general value" for rook. The value exists for the sake of our calculation and I think we are on the same page on this as well? 

So ultimately we don't need to fight over whether bishop is slightly better or not. It's too insignificant for us. As long as we have a general understanding, that's all that matters. 

Yeah I can agree with that for a knight vs a bishop, but like I said earlier I think master practice favors the 2 bishops over bishop and knight or 2 knights.

Thought experiment: you conk your head and get a specific amnesia where you remember everything about chess except piece valuations. How much worse would you play?

I think I would lose 500 ratings points.