Basically alphagozero just know the rules. It doesn’t use human knowledge of the game at all.
Then in a few days by training he can go ahead of what humans researched for thousands of years and even beat them
CivVI is a different game, different rules doesn’t mean it’s so complicated. First you can dismantle in sets which can use different way of AIs (like heuristic) not only the last updated from Nature. To me this game doesn’t seem to have a combinatoric as deep as Go. (Number of atoms in the universe, ya know that I guess) look at chess: unlike Go, you have different pieces, they move… That doesn’t make chess so complex compared to go
I may be wrong ofc
Coming back on the civ series, I did enjoy a civ V well modded (with simfanatic offers if my memory is good)They made marvels combining civ IV in it. So I push everyone to try it.
I’m somewhere in the middle on how effective that sort of machine learning would be for Civ VI. I think the bar is so low it couldn’t fail to produce a much stronger AI than the current one, though. :)
It’s worth noting that it wouldn’t take just a few days to get there. Playing against itself, it was able to play hundreds or thousands of games of go a day. Civ VI isn’t as well optimised for playing at that sort of speed.
Also, “just knowing the rules” wouldn’t be as simple. There are all sorts of bugs, different ways of rounding, minutiae of resolution order, that sort of thing that mean I’m pretty sure even the devs couldn’t write down the rules of the game without a significant amount of effort. You’d certainly end up with an AI that exploited every bug to the max.
I gave you a very concrete list of differences which aren’t trivial. I also listed three things that were specified by hand, which you only learn about if you read the actual papers not the promotional blog kool-aid.
You just did some jazz hands at me.
What you said @rho21. A big issue using Monte Carlo Tree Search is the sheer computational cost of constructing the tree of game positions and moves putting those together. It took 40 days to train AlphaZero and that on a custom made super computer of which as far as I know there’s exactly one in service (Deepmid’s).
OpenAI had an, arguably, easier time with DOTA, accessing the game state via a programming interface, and constructing a simplified yet informative representation of game statwz. That was used by the AI players that got hyped a few months back. But OpenAI weren’t doing as Deepmind’s AphaZero, that plans moves ahead, they were just training a NN that predicted the payoffs of the games moves.
Regarding fitting to bugs etc. That has happened already and quite often. Deepmind Atari player broke the game of QBert, exploiting a bug that lied dormant for 30 years.
You may be interested in the, quite amusing, examples compiled on this wonderful blog
I did read carefully your concrete list and did answer giving you 1 famous example of something which looked more complex (chess) but which is not.
To say that the Nature paper is a lie ( by omission) sounds a bit strange in a research community. They even got a AI prize I think for their work? I did just report what they wrote: no human knowledge on the game. Is it a promotional lie? In fact I am really interested in that concern so if you could explain a bit more your opinion, I would read it carefully too.
Please don’t take offense as this is absolutely not my intention.
I myself was one of those people who would never believe it could happen or at least not before some decades. And the second shock was when they made this “no human knowledge” version which is even stronger.
Concerning the use of a super computer, that’s true but now you can run on your own PC (better with a good GPU) a lighter version which is still quite good. Alphagozero did enter everyday life of many go players (even professionals of the game) as go assistant. And that could interest the civ players of the future.
There are interesting points in the debate around this AI aspect I found, like to make the game more compatible to the AI as the AI to the game, or how the AI would maybe first more debug the game mechanics. I’m far away to think it would be a task too complex to not enjoy it in our civ games, because you know I am getting cautious on this kind of prediction in which I went wrong twice already.
Chess was “solved” 20 years ago - remember Deep Blue vs Garry Kasparov? Not sure yet what was the point of that. Especially since chess hasn’t any of the features I listed.
Who says that? The paper is fine. You can find all the nitty gritty details there.
The blog post you linked is a different matter. Is hyping the same as lying? Americans have the expression “astroturfing” for that.
I listened again to Soren Johnson’s Designer Notes episodes with Brian Reynolds and @Sid_Meier (dang, that @ is the only one that wouldn’t work), and they discuss regularly the city spam issue of earlier Civs and the long road to mitigate it… I don’t know if they ever succeeded — I am anything but a Civ power player — but it seems, reading you, that Civ 6 re-opened Pandora’s box!
Pardon my useless venting, but I need to type it: Civ 6’s AI is so infuriatingly bad it turned pacifist-baby-culture-victory-me into warmonger-bloodthirsty-new-me then, and in all the strategy games I have played since. I feel I am still angry at it.
Trying to balance Tall versus Wide empires is always an interesting challenge since ideally both would be viable strategies in the game but it certainly doesn’t always work out that way. From what I remember of Civilization V multiplayer the power play was to work your way up to four well developed cities. So Civilization V leaned towards the other end of the spectrum and typically universally favoured taller empires over wider empires.
Haven’t sunk enough hours into Civilization VI to know much about what strategies are considered optimal and such. Wouldn’t surprise me if Firaxis accidentally skewed things back towards wider empires being the preferred strategy in response to Civilization V.
That’s the article the bit of text I quoted came from. Sulla produces some excellent reports and dissections of Civ and other games; almost invariably accurate and incisive. His excellent preview walkthrough of Civ IV is what drew me into playing the game seriously.
Sulla actually was quite positive toward Civ VI - with the exception of the AI. Civ V has perverse rules regarding expansion that Civ VI lacks, but has a (marginally) better AI.
He ultimately seems to have abandoned Civ VI singleplayer because of that AI.
That is from before the two expansions were released though, so while the base game started out favouring wide empires Firaxis over the course of the game’s lifespan eventually swung the balance the other way. Once the Brave New World expansion rolled around you’ll see that even Sulla says that the pattern of three to four cities is the dominant strategy in their Brave New World review.
So by the end of Civilization V’s lifespan, which I’m going to have the freshest memory due to recency and playing the most, Civilization V does indeed favour Tall empires over Wide empires. A lot of Firaxis’ thoughts for Civilization VI were in reflection to the state of Civilization V post-Brave New World.
I think Tall or Wide will always be mathematically superior to min/maxers, I just want both strategies to A) be viable and B) offer different approaches to the game. If Wide is better, that’s fine. I just want a decent shot at beating the game going Tall (or vice versa).
It’s strange reading some of Sulla’s reviews where i both agree with everything he has written in laborious detail, but get the feeling he misses the point at the same time, because he likes Civ VI and i felt that Civ VI was even worse trash than Civ V’s many combined problems, and that he seems so right about Civ Vs but so forgiving of Civ VI, like the game breaking free resource rush you get for finding city states first and the completely, utterly useless, dead ended rabbit hole that is Faith, that it makes me question if he really has the big picture… or what, exactly, that big picture really is.
There’s more to a strategy game than some devil’s dance between super deep hardcore min/maxing aspie edgelord and noobles pushing pieces around the game board grunting of pleasure as they stuff their pieholes with cheetos and press Next Turn. Call it the “Third Way of Understanding Strategy Games” maybe. I’m not 100% sure i can exactly define it, but i know it when i see it.
It’s a common problem I’ve noticed among people who tend to obsess over every tiny number in a particular system. They sometimes overlook how all the systems interact with each other as a whole, or the conclusions are only true when looked from one narrow playstyle.
Agree completely. I think that problems kick in, though, for those who play primarily multiplayer.
In truth, ANY strategy can compete successfully in single player, as long as you make some plan for dealing with an early attack. Certainly, both tall and wide are entirely viable.
But the crap AI pushes a lot of people (not me) into multiplayer, and some of them become a bit over the top about balancing everything – which is, in itself, not a recipe for good gameplay. (I have always felt that everything from civs to governors would be more fun, if somewhat less attention were paid to balance.)
Very true! It’s why I don’t tend to play strategy games competitively in MP, because ironically it tends to eliminate 90% of potential strategies one can take.
But I think some very basic things alter players’ reactions to a game in a very deep way. One is that a game like Civ is an extremely different subjective experience the first time or two you play through it, as opposed to the 20th or 100th time.
It is also different depending on whether you are driven to use a game wrecking exploit or not… Whether you are happy playing until it gets boring, and then starting over – or whether you just have to see that victory screen… Whether you have to know the AI opponents are playing by the same rules as you or whether you can ignore AI cheats… and so on.
I have not read Sulla’s reviews, but it sounds like my heart is in the same place as his. I could fill a page listing obvious grievous faults with Civ VI, yet, at the end of the day, I know I have gotten more enjoyment out of the game in the past couple years than all other games combined. On the other hand, I got very little enjoyment out of Civ V. However, my objective analysis of the two games would not clearly support that.
Perhaps because the first few games of Civ VI were so bad to me that i never bothered to really learn the systems, but unlike Civ V or any other Civ game for that matter(I liked Civ III’s perhaps the best despite having obvious flaws) i don’t really “get” Civ VIs convoluted and opaque underlying systems very well. I’m not sure about amenities; i’m not sure about housing; i’m not sure about happiness. I mean settle near rivers because i get +2 housing or something… sure. Wonders often seem under-powered if not actually superseded by many regular buildings.
Yet Tomara gets two for the price of one, so i just run around and whomp everything and none of that stuff matters. Or i make a bunch of roads with Rome and whomp everything, and none of that stuff matters. I’m bothered Civ VI has a bunch of layers to its onion that i don’t seem to even have to understand to play “well”, and then it underestimates or overestimates the value of many of its systems relative to the scope of the game (or so it appears to me). Like, it seems like a good 1/2 of districts will never, ever be built if you min/max the game, for example.
Of course i also think God King in Civ V is by far the best economic Pantheon you can get in almost all cases! #fightme.
Agreed. While I played a lot of multiplayer it was often with friends, usually I wasn’t overly concerned with min-maxing the game through the roof. I just had a general strategy in mind and then broadly went along with it rather than getting bogged down in the minutiae. So I just want a Civilization game to broadly provide some meaningful strategic\tactical choices and a challenge that hits that “enjoyable” spot.
Funnily enough, my preferred Brave New World playstyle just happened to coincidentally fall in line with what the “top 1%” considered optimal rather than being deliberately min-maxed.
