A couple of years ago there was a storm of coverage that came to the newly introduced AI neural network known as alpha zero, the ai dominated a lot of pursuits.
What I will be talking about today is the accomplishments of Alpha zero when it comes to chess (since this is a chess website) and how is it stronger than any other traditional engines.
But first, what exactly is alpha zero?
AlphaZero is a computer program developed by artificial intelligence research company DeepMind to show the potential of a.i. neural networks by mastering chess, shogi, go, starcraft, and protein folding.
Can Alphazero be beaten?
It’s not possible for any chess grandmaster to beat Alphazero. Even the strongest traditional chess engine stockfish 8 was defeated in 100 games (Out of 100 games 28 were won by Alphazero and 72 draws). However, newer versions of Stockfish may defeat Alphazero (Stockfish 12).
Is alphazero the strongest chess engine?
Deepmind’s Alphazero managed to defeat Stockfish 8 in their latest clash of 1,000 games where 155 games were won by Alphazero, 6 games won by stockfish, and 849 games were drawn. Stockfish has improved since and it is still a question if Alphazero can overcome Stockfish 12.
I just want this article to be a whole discussion about the strength of alphazero and how it is different from other traditional chess engines. I think you will be interested in this in particular since it almost plays like a human even being an a.i.
I sure am interested, let’s begin.
Algorithm of alphazero vs. traditional chess engines
In order for alphazero to stand out, it has to follow a unique set of algorithmic choices that are different from a traditional chess engine.
After all, the reason why so many engines have similar strengths is that they are running the same methodology when finding their moves (meaning they are not that different from one another).
Alphazero presents various methods internally when finding its own sets of moves.
Alphazero follows a monte carlo search tree
This has been the key difference that separates this engine from traditional ones, which is the way it tries to find the moves (in a search tree).
Alphazero follows what’s known as a Monte Carlo tree search, a different search tree algorithm than stockfish that analyzes fewer positions but is claimed to be more efficient.
Alphazero only analyzes 60,000 positions in a second, while stockfish analyzes 60-70 million positions in the same second.
A search tree is basically a conceptual framework that analyzes the continuations of each move and does it until a stable position has been confirmed.
Now, you would think that the more that an engine can search in a search tree, the more that they are able to find stronger moves.
And traditionally this is the case, however alphazero shows that being more efficient (focusing on a set of moves that are most reasonable) provides deeper evaluations on the lines.
This allows it to focus on those that actually matter and try to dig deeper on those (not necessarily considering so many moves all at once).
Alphazero does not rely on brute force calculation
Brute force calculation is simply a term to express a methodology that outcompetes an opponent by finding as many potential moves as possible. This is also the reason why a human finds it very difficult to beat a traditional chess engine.
The traditional chess engine has more capability to find broader options when it comes to moves than a regular human (humans have to think of other things too) which is why it is called beating an opponent by “brute force calculation”.
Alphazero is different though, unlike Stockfish who relies on brute force calculation (evaluating millions of lines), Alphazero has been simply put, works smarter not harder (which also gives it a unique perspective on the positions).
By only considering moves that actually matter, it is more capable to explore those lines in longer continuities.
Stockfish may not consider a move that has analyzed certain levels deep because the endgame might not be as favorable, but Alphazero can see beyond those since it doesn’t have many options to consider.
In games between Alphazero and Stockfish, Alphazero makes moves that Stockfish would not even consider at first, but will actually recommend it if the traditional engine was allowed to analyze the position for hours (and with a greater depth).
This proves traditional engines can at least make sense of Alphazero’s reasoning when it comes to moves, but would have to do so in greater depth.
Alphazero follows an artificial neural network design
When it comes to traditional chess engines the methodology is like that of a machine, where there is a set of systems that finds the moves. Alphazero however follows a methodology similar to our own neural network.
Unlike with Stockfish and all the other famous chess engines out there, Alphazero is modeled after an ANN (Artificial neural network) or simply called NN (neural network), which are computing systems vaguely inspired by the biological neural networks that constitute animal brains.
The way to search out the moves are done in a similar fashion as an animal brain than just by relying on a simple search tree (though Alphazero also uses a search tree).
This is why alpha zero plays in a human-like manner, the way it makes decisions is similar to how humans make their own decisions.
But you might think that alpha zero would just get crushed since actual human players are not capable of beating traditional chess engines.
But that is actually wrong, real humans have other things to take care of outside of chess, meaning we don’t really get to use the capacity of our brains to its full potential.
Alphazero takes our neural function to the next level and only focuses on a particular task (chess in this example), think of this as the epitome of human potential.
Style of play: Alphazero vs. traditional chess engines
Yeah sure, alphazero is different since its internal algorithm is arguably superior to modern chess computers. However I want to discuss how Alphazero differs from these other engines when it comes to their style of play.
This can give us a picture of how to approach chess at its highest level, and how alphazero brings a lot of interesting things to the table.
Alphazero prioritizes piece activity
When we were first introduced to chess, most of us are taught that piece activity matters more than any other aspect out there, even over material advantage.
Though traditional engines implement this same idea, it doesn’t execute this concept to the level alphazero does.
In the seemingly drawn positions that are given to stockfish for example (In Stockfish vs Alphazero), it has been seen that Stockfish goes for material in order to break equality.
In the games that exist between Alphazero and Stockfish, we can observe that Alphazero prioritizes piece activity more than material advantage (it would even give up material occasionally in order to activate its pieces).
And this is during positions that are seemingly drawn!
It has a tendency to sacrifice multiple pawns if it meant activating a dormant piece that can help break a stale situation.
Occasionally stockfish would give an evaluation to these sacrifices as being an advantage for their particular color, but it still ends up losing later down the road.
Although engines do prioritize piece activity if it is the best move in their algorithm, alphazero just does it on another level.
Alphazero is semi-aggressive
When we look at games played by traditional engines, we can see certain decisions do gravitate more on drawn positions, it is unlike a human where games are much more likely to be decisive.
This is primarily because traditional engines mostly play safe lines unless a reasonable attack would bear fruit a couple of moves later. Alphazero on the other hand seems to have more liking in placing its pieces in a way that attacks will naturally flourish.
The style of play Alphazero offers is just quite different, it is more aggressive and would even consider unheard sacrifices in order to start an attack.
Traditional engines do execute sacrifices as well, but are much rarer on the occasion that they are playing against another engine.
Alphazero on the other hand attempts attacking ideas even if the combinations aren’t there yet, it makes its move even if there is no apparent advantage in doing so. Some will argue that this is a weakness but it is apparently working in this case.
Coupled with the efficiency of only analyzing fewer lines with greater perspective, this attacking tendency might pay off.
And it did pay off in the couple of instances alphazero is playing against other engines, this semi-aggressive quirk might be one of this neural network’s greatest strengths.
Alphazero does a lot of squeeze play
In chess there is such a thing as a squeeze play, basically doing certain decisions that restrict the opponent’s ability to make their pieces more active. Again, traditional chess engines are also capable of doing this but Alphazero seems to be more capable.
Alphazero’s algorithm suggests squeeze play, greater board room, and prevention of development. In the games with Stockfish, it can be noticeably seen that Stockfish always struggle with space (being too narrow to have room for any play).
It can be seen that alpha zero would even sacrifice a couple of pawns in order to restrict some pieces of stockfish. It even has a zugzwang immortal against stockfish itself by doing this exact principle.
This just further proves its superior strength when it comes to squeeze play and not only that it executes it, but it knows how to take advantage of such positions.
Engines such as Stockfish could theoretically thrive even in these scenarios, yet alphazero clearly has an edge having an overwhelming result when there is a specific situation that a squeeze play is present (even at the cost of material).
Alphazero values long-term compensation
Long-term compensation refers to having an advantage as a result of a sacrifice for example, yet is not immediately seen since it still has to be played out. There’s no visible combination that is apparent right off the bat.
Long term compensation is a hard concept to grasp for human players, and even if traditional chess engines have it too (sense of long-term compensation), Alphazero seems to be much more sensitive to it.
Sacrificing several pawns in order to play an endgame where one of their bishops, knight, or rook is much more active is something that I have seen alphazero is capable of deciding. Other engines have that but not to the extreme extent that alpha zero does.
The sense of long-term compensation is sometimes so unconventional that a lot of other chess engines do not agree that it is the best idea, yet alphazero was able to convert anyway.
Chess is not only finding a definite advantage based on our own set of standards (theoretical value, endgame combinations, etc.) but the actual richness of the position.
Some positions are just so rich that alphazero was able to execute unconventional long-term initiatives.
Alphazero is more likely to avoid draws
One thing that separates a human from an engine is the desire to actually win, chess engine just evaluates the position and finds the best moves.
This means that if a position is drawn, an engine will continuously play for a draw since it is just the best option.
There are certain instances in human play though that a player would still push for initiatives in order to gain some sort of an edge, even if putting up some risk by doing so.
I think this is exactly the nature that allows alphazero to have an upper hand against traditional chess engines.
Alphazero tends to gravitate toward avoiding draws than other traditional chess engines, it would try to do some unusual ideas in order for some chances to arise.
This does not mean that playing for a win when a draw is possible is always the best idea, but it does allow alphazero to grab some decisive results from seemingly drawn positions.
Alphazero has bits of creativity and innovation
With so many theories in chess it is easy to assume that the game is nearly solved, that there aren’t any gray areas left to explore. However, alphazero learns to play chess by itself away from the potential influence of modern theoretical knowledge.
I think this is the reason why Alphazero occasionally considers moves that are not even suggested by the top engines (and vice versa), which prompts the idea that it is quite innovative.
It is actually the engine that tainted the reputation of the one’s famous french defense and queen’s indian defense.
The French defense and Queen’s Indian defense are pretty established theoretical openings that even modern chess engines do not consider to be suboptimal, yet alphazero has destroyed it.
This tendency to not follow conventional accepted ideas may be one of this neural network’s strengths.
Alphazero learning on its own makes it groundbreaking
The reason why as to alphazero is named that way is basically that it has learned on its own without access to modern information.
Alphazero basically means starting from zero experience, it taught itself chess by playing against itself without outside influence.
In analogical terms AlphaZero is like a brilliant person, who does not have any teacher or mentor; it looks at Chess in its own way and gets better at it.
Groundbreaking entities are usually those that are not fully aware of major knowledge in a particular field, they are just much more likely to find unexplored territories that a lot of other people/engines will overlook.
I just have to mention this again, I think this is one of the reasons that Alphazero has thrived against other strong chess engines. It is just able to explore territories that are being limited to the other traditional engines algorithm-wise.
Alphazero is better at time management
Few of you might know this but chess engines do also need time to process information, they don’t evaluate positions in milliseconds all the time.
It can evaluate that fast but usually not on its maximum depth capacity, in order to be more accurate it needs time to go look longer on the appropriate lines.
This is why during battles of chess engines there is such a thing as time control, and if an engine takes too much time to think of its moves it will be officially resigning (time qualifies as one of the signs of a resignation).
Alphazero seems to not only be overall better, but is being better and doing it in a faster rate.
Alphazero seems to be better at dealing with time constraints than other chess engines.
When AlphaZero Crushes Stockfish In the New 1,000-Game Match (link), AlphaZero bested Stockfish in a series of time-odds matches, soundly beating the traditional engine even at time odds of 10 to one.
This is the epitome of efficiency, by not needing to look at millions of lines it can find the best moves in the shortest amount of time possible.
Alphazero’s style feels like that of a human
The thing with engines is they implement a brute force calculation, the things that they do are usually things that cannot be understood by human viewers.
Alphazero though is different, it follows more objective reasoning that humans could at least fathom on a surface level.
It plays more like a human, it has tendencies to make decisions that do not make sense algorithm-wise but would be seen as reasonable by a decent human player.
I think it is because of the lack of brute force calculation, it doesn’t look at so many lines all at once that it plays moves humans can’t consider.
It only prioritizes the most reasonable options therefore having a repertoire that can objectively be understood by human-level understanding.
Plus it runs on a neural network-based algorithm that is based on animal brains, brains that we naturally possess. This means that all the moves that it considers (although not all the time) can at least make sense from a human perspective.
Can alphazero beat the latest stockfish?
There has been a whole fight online debating about the chances of stockfish with its new neural network upgrade (stockfish 12) if a rematch against alphazero occurs. Now, I know that I am not supposed to be here to just say opinions but we just don’t know.
It’s a question whether Alphazero can achieve the same feature against Stockfish 12 and above (who also has a neural network now) but it would be very interesting to see.
I personally believe that stockfish 12 is now capable of beating the alphazero that we have seen before, unless alphazero himself has experienced an upgrade.
The developer of stockfish has now patched the flaws that were exposed by alphazero and is hailing as the top engine of the world. Plus the fact that it also now has a neural network design makes it a very formidable opponent for our neural network alphazero.
But alphazero has only spent about 4 hours or so trying to learn chess before, so the ceiling of alpha must be higher.
I really don’t know if you are looking for evidence, but one thing I am sure of is it will be a firework if it were to happen next time! (a rematch that is)
Do you now know why Alphazero is so strong?
The development of neural networks in every industry not just in chess is fascinating to me, we really are indeed in a revolution. I am so happy to be alive at this moment to see how far we can push this concept, and how it can make the world better.
I really wish Alphazero was made available for the public in order for us to test its full potential, and pit it against every engine that we want. It is unfortunately not the case and this mystery might never be solved if Deepmind doesn’t push this through.
However the possibility of neural networks will never cease even if it is not about chess. I am hoping to see it, sleep well and play chess.