Could have just posted this under the 'The End is Near' thread. . . .

What about Starcraft?

Never played it!

Quote:

---

Algorithmic optimization is better done by computers than by humans in complex closed systems.

---

I'm not sure it actually shows this. Could be that human understanding of Chess is lacking.

One of the interesting things about this experiment to me was that the machine considered several openings, but ultimately landed on human ones.

I would like to see the results of this when it is limited to maybe looking at 5-10 positions a second. This is probably more in the human range. I wonder what it's performance would be like at that point.

dargscisyhp said:

---

Quote:

---

Algorithmic optimization is better done by computers than by humans in complex closed systems.

---

One of the interesting things about this experiment to me was that the machine considered several openings, but ultimately landed on human ones.

---

I haven't searched for an article, but wanted to understand what you meant by this?

Do you mean that the "Alpha Zero" did self practice and came to agree with the openings that humans use or that it was already aware (or googled them ) and utilized them as optimal to open with?

The first option would be a really cool finding.

dargscisyhp said:

---

Quote:

---

Algorithmic optimization is better done by computers than by humans in complex closed systems.

---

I'm not sure it actually shows this. Could be that human understanding of Chess is lacking.

One of the interesting things about this experiment to me was that the machine considered several openings, but ultimately landed on human ones.

I would like to see the results of this when it is limited to maybe looking at 5-10 positions a second. This is probably more in the human range. I wonder what it's performance would be like at that point.

---

I wonder if you could do the reverse experiment and crowdsource moves from the 1000 best human players for a match, thereby upping the human computation power.

To your other point about us maybe not understanding chess, all the top players practice against computers they can't beat. So maybe we are learning something.

amercer said:

---

dargscisyhp said:

---

Quote:

---

Algorithmic optimization is better done by computers than by humans in complex closed systems.

---

I'm not sure it actually shows this. Could be that human understanding of Chess is lacking.

One of the interesting things about this experiment to me was that the machine considered several openings, but ultimately landed on human ones.

I would like to see the results of this when it is limited to maybe looking at 5-10 positions a second. This is probably more in the human range. I wonder what it's performance would be like at that point.

---

I wonder if you could do the reverse experiment and crowdsource moves from the 1000 best human players for a match, thereby upping the human computation power.

To your other point about us maybe not understanding chess, all the top players practice against computers they can't beat. So maybe we are learning something.

---

I doubt if you put all the top human players together they would achieve much better results than the top human. It just doesn't seem to scale that way. They have crowdsourced games on chessgames.com against strong players, and I don't think they typically outperform the person they're playing. Mind you many of these amateurs who are putting their minds together are extremely strong in their own right, and occasionally another professional will lead the crowdsourced team. The advantage humans have is in how they prune branches from the game tree. We will probably have that same analysis done 1000 times, therefore quite a bit of that extra computing power will go into redundant calculations/evaluations.

To your second point, while modern engines not named alpha zero are great for checking your ideas for outright errors, they have zero understanding of Chess aside from what we programmed them with. While they can probably show errors in our calculation, and maybe give us some heuristics to play by based on their calculations in extremely complex positions, they're not going to teach us much about Chess. With alpha zero, already one of the things that might be a revolutionary idea in Chess is that humans might currently overvalue pawns. It was dropping them left and right for more abstract ideas like having the initiative. This is one of the things that we can glean just from having ten publicly available games from it.

kurt vonnegut said:

---

Could have just posted this under the 'The End is Near' thread. . . .

---

More birthpangs.

Repent!

AstroAg17 said:

---

It's the first scenario. Alpha Zero during its learning phase experimented with pretty much all of the most popular openings. Here are graphs of which openings fell into and out of its favor over time.

---

Gotcha. So it was aware of the popular openings, but didn't have to use them. During it's learning it figured out that they were in fact the "best" openings.

That's interesting on a couple fronts because it seems to validate that humans are on the right track with our openings, but that we make sufficient mistakes on a completely transparent game to lose anyways.

Did it find any openings that we underutilize or didn't know about as effective openings?

Ulrich said:

---

I like to think that this validates my carelessness with pawns and frenetic style.

---

Oh, definitely, me too. We're just ahead of our time.

AgLiving06 said:

---

AstroAg17 said:

---

It's the first scenario. Alpha Zero during its learning phase experimented with pretty much all of the most popular openings. Here are graphs of which openings fell into and out of its favor over time.

---

Gotcha. So it was aware of the popular openings, but didn't have to use them. During it's learning it figured out that they were in fact the "best" openings.

That's interesting on a couple fronts because it seems to validate that humans are on the right track with our openings, but that we make sufficient mistakes on a completely transparent game to lose anyways.

---

As Astro has said, it was not aware of any openings beforehand. However, in the course of its learning it did stumble onto our openings, and analyzed them to quite a degree it seems. Also, the board is anything but transparent. It's hidden behind the wall of combinatorics. There are more possible board arrangements than atoms in the universe.

Aggrad08 said:

---

Did it find any openings that we underutilize or didn't know about as effective openings?

---

No, but I do somewhat wonder if it invalidated some of those lines which it stopped playing.

dargscisyhp said:

---

AgLiving06 said:

---

AstroAg17 said:

---

It's the first scenario. Alpha Zero during its learning phase experimented with pretty much all of the most popular openings. Here are graphs of which openings fell into and out of its favor over time.

---

Gotcha. So it was aware of the popular openings, but didn't have to use them. During it's learning it figured out that they were in fact the "best" openings.

That's interesting on a couple fronts because it seems to validate that humans are on the right track with our openings, but that we make sufficient mistakes on a completely transparent game to lose anyways.

---

As Astro has said, it was not aware of any openings beforehand. However, in the course of its learning it did stumble onto our openings, and analyzed them to quite a degree it seems. Also, the board is anything but transparent. It's hidden behind the wall of combinatorics. There are more possible board arrangements than atoms in the universe.

---

What I meant about transparency is that there is no hidden information. What's on the board is the available information, nothing is hidden.

Kasparov's new book is on A.I. This new learning process may require an update.

This may be the end of Nick Saban and Bill Billicek.

Couldn't you enter the rules of football and generate never before seen plays, formations, and strategies? Crazy thought.

Wede01 said:

---

This may be the end of Nick Saban and Bill Billicek.

Couldn't you enter the rules of football and generate never before seen plays, formations, and strategies? Crazy thought.

---

Football would be a lot more complex. It's not a perfect information game, there are a lot more factors to consider (players' particular skillsets as opposed to set in stone rules for piece moves, weather where no analogue in Chess exists, perhaps physics itself to be able to simulate a game etc.) and chance almost certainly plays a part. For these reasons, I imagine we are a long way away from that.

dargscisyhp said:

---

Wede01 said:

---

This may be the end of Nick Saban and Bill Billicek.

Couldn't you enter the rules of football and generate never before seen plays, formations, and strategies? Crazy thought.

---

Football would be a lot more complex. It's not a perfect information game, there are a lot more factors to consider (players' particular skillsets as opposed to set in stone rules for piece moves, weather where no analogue in Chess exists, perhaps physics itself to be able to simulate a game etc.) and chance almost certainly plays a part. For these reasons, I imagine we are a long way away from that.

---

Not to mention execution. You can't gameplan for fumbles, dropped passes, overthrown balls, injuries, etc.

Good points, I was trying to lazily chalk all that up to "chance."

If you actively play Chess, this is an interesting watch.

Thanks for that, I enjoyed it, in a scared humanoid kind of way. Who may I ask was the narrator?

He's NM Jerald Meyers, I believe.

An open source attempt to replicate the results of alpha zero is under way. You can donate your coy time if you want. You can even play it.

https://www.reddit.com/r/chess/comments/882jr8/leela_chess_zero_the_fork_of_leela_zero_for_chess/