AI beats top human player. Yet another AI discussion

[Disgustipated]

http://money.cnn.com/2017/08/12/technology/future/elon-musk-ai-dota-2/index.html

Interesting in that an AI beat the best player in the world. So the question is, how difficult do you want your AI? I think most players would be frustrated if the AI was too good. It's why I don't think Firaxis will ever invest serious resources in it. And honestly, I don't want an AI that great either. But it would be nice to have an AI that didn't make embarrassing boneheaded mistakes an 8 year old would never make. It doesn't feel like much of an accomplishment when you beat such a dumb AI.

Of course take note that the AI in this article required massive computing power, but it's an interesting article, I suggest reading it. I'm not familiar with the game in question, so maybe someone can fill me in.

Ideally I'd love for all the handicaps to be even (Prince level) and you have a slighter that can dial up to increase the decision making process of the computer players. We are years if not decades away from such a thing, and with Firaxis' limited budget, it probably won't happen for this franchise. Your thoughts?

"Our bot is trained entirely through self play. It starts out completely random with no knowledge of the world, and simply plays against a copy of itself -- which means it always has an evenly matched opponent," Brockman said.

The bot kept playing until its skill level reached that of the world's best Dota 2 players, he added.

 

[Archon_Wing]

Three things:

I didn't think Dota 2 was a 1v1 kinda thing. 1v1 also isn't your average Civ game either.
Dota 2 is an esport. Thus players of the highest caliber cannot be challenged by your average consumer AI , and would probably only partake against the best AI there possibly is.
AIs have an inherent superiority in games that require accuracy, quickness and speed.

So indeed, I wouldn't expect anyone to invest in such a thing, because it would only appeal to hardcore players that would rather play each other anyways.

It is a hard question to answer, because as I've said elsewhere, your average player doesn't come on forums or reddit to discuss game mechanics and thus it's entirely possible the stock AI is already enough for them. I would expect the base AI to at least be able to master mechanics and be superior to a player that hasn't though. However, realistically speaking the AI will need bonuses to keep up with players with a clue though currently it is a lot due to the badness of the AI. The huge bonuses are evident on how confident they are on it-- not very!

Part of what I see in Civ 6's AI is the idea of "knowing how to play". In the strategical context, this would seem a bit limited by the knowledge of the designers. If the designer of the AI does not understand how the game works, then the AI suffers accordingly. Now, consider the AI is done before release and often before the majority of players know how to play the game. And then consider that collectively speaking after release, the combined knowledge of the players will always quickly outpace the knowledge of the developers. Therefore, it is often mandatory for the developers to listen to the players on how the meta works. Problem is, who do you listen to for this? Unfortunately, the loudest ones aren't often the most skilled. And we don't have a ladder do we? There are many, many cases, where bad ill-informed players have pushed for game design choices that are bad.

 

[elitetroops]

Ideally I'd love for all the handicaps to be even (Prince level) and you have a slighter that can dial up to increase the decision making process of the computer players.

The AI needs handicaps in Civ, because the AI also needs to be forced to make suboptimal decisions. If the AI did only optimal decisions, like it can do in a game of Chess or Go (I don't know what Dota 2 is like), then every AI leader would act the same and the whole concept of diplomacy would be thrown out the window. As long as different AI leaders are supposed to have different personalities, and to treat other leaders differently based on their diplomatic relations, the human player has a huge edge in being able to do whatever he/she wants at any time.

 

[stinkubus]

Have any ambitious programmers ever taken a crack at programming a Civ playing AI?

 

[UWHabs]

I don't want the AI to crush me every time, but yes, I would appreciate an AI that simply "didn't make mistakes". It also sucks the level of bonuses they get - I mean, when they start with tons of units and multiple settlers on higher levels, it's not rare for an AI to potentially have 4 cities in place before I get my 2nd (2 from their initial settlers, 1 captured from a city-state, and then 1 settler on their own). I'd rather they get more bonuses through the game as opposed to so many bonuses early on, since it can be frustrating starting a new game and seeing them jump so far in front.

 

[miaasma]

as someone who actually plays dota 2 (unfortunately) it's a little more complicated than "an AI beat a human at dota"

if you got five AI to beat five professional dota 2 players it'd be more of a statement, but since 1v1 games of dota mostly come down to a certain rhythm of behaviours it actually makes sense than an AI would become more proficient than a human. it's mostly muscle memory and reactions, both of which a computer will always be better than a person at

 

[Zuizgond]

The AI needs handicaps in Civ, because the AI also needs to be forced to make suboptimal decisions. If the AI did only optimal decisions, like it can do in a game of Chess or Go (I don't know what Dota 2 is like), then every AI leader would act the same and the whole concept of diplomacy would be thrown out the window. As long as different AI leaders are supposed to have different personalities, and to treat other leaders differently based on their diplomatic relations, the human player has a huge edge in being able to do whatever he/she wants at any time.

I prefer the AI to never make suboptimal decisions personally.

 

[stinkubus]

If that were the case the game would be unwinnable and turns would probably take days in the end game.

 

[Magma_Dragoon]

This isn't very impressive. An AI can keep track of respawn time of jungle creeps and optimize its itinerary flawlessly while in the middle of a fight, it can land a last hit or deny every time, predict perfectly how long it will take courier to deliver a newly purchased item and make plans based on when it will get it, it can calculate based on it's last visual contact of the other player the exact limits of where the player could be, predict the last frame it can use blink dagger and use it accordingly, etc...

as someone who actually plays dota 2 (unfortunately)

When it's good it is very fun, when it's bad it is the worst thing you can do to yourself without drawing blood

 

[bbbt]

As noted, beating a human on a real-time game is not quite the challenge as turn-based (or with as many variables as civ!)

Also, dota being basically online and having many players (and not much in the way of mods afaik) and fewer variables makes it much more possible to use machine learning on than civ.

Realistically, no most people don't want a very challenging AI in civ, I've said before the vast majority of players play at prince or below based on the achievement stats.

Having said that, Civ does have one advantage for being more challenging than something like Dota AI-wise, in that you can theoretically have many competitors on the map - so as long as it's challenging enough that one can win while you are dealing with others, you can lose.

So if the AI gets to the ability to be somewhat competent AND have a variety of playstyles it's probably enough - i.e. the one that turtles and just expands/defends and beelines will win while you are dealing with the over-extended competitor.

My ideal CIV AI set-up would actually be something like 4 sliders:
AI bonuses
AI 'unprovoked' agressivness
AI victory beeline (i.e. how early it starts aiming for a victory versus just building a 'balanced' empire)
AI competitiveness/envy (i.e. how much it dislikes and tries to stop you when you are ahead, versus just roleplaying and doing it's own thing).

And adjusting those 4 sliders would let you pick the AI 'experience' you prefer for the game.

 

[Ryika]

I think the main problem with assessing a decent level of AI difficulty is that the "difficulty" of playing Civ does not come from playing against AIs, it comes from playing the "map", and making optimal choices to increase the efficiency of your empire. AIs are just entities within the same game, that you will generally not be confronted with, aside from that one guy that you'll be at war with at any point in time.

All other Civs play parallel to each other most of the time, so having an AI that is just very good at that, leaves the player with little room to understand what they're doing wrong, and why the AI can outperform them.

And the problem of runaway-AIs that you just can't deal with if they start in a secluded position.

 

[Acken]

The AI needs handicaps in Civ, because the AI also needs to be forced to make suboptimal decisions. If the AI did only optimal decisions, like it can do in a game of Chess or Go (I don't know what Dota 2 is like), then every AI leader would act the same and the whole concept of diplomacy would be thrown out the window. As long as different AI leaders are supposed to have different personalities, and to treat other leaders differently based on their diplomatic relations, the human player has a huge edge in being able to do whatever he/she wants at any time.

Which is why their individual flavorful choices have to be balanced to begin with if we want both diversity and the AI to not be awful.

 

[sonicmyst]

AI in Civ needs to be competitive but still make the game enjoyable. A too difficult AI will mean frustrated player.

 

[Magma_Dragoon]

AI in Civ needs to be competitive but still make the game enjoyable. A too difficult AI will mean frustrated player.

Those players could improve by watching the AI and copying how it plays

 

[stinkubus]

Any top shelf AI would certainly have randomizing elements (because some features of the game are random) and it would take hours upon hours of careful observation just to be certain you even knew how it played.

 

[Raz.BELOVED]

As someone who avidly plays dota 2, Dendi is not a top player. Him and his team did not make it to the group stage of the biggest dota tournament in the world (TI) which ended on saturday. He was previously a top player, and is still the biggest household dota player name because of his personality and performance in the 1st, 2nd, and 3rd TI, but this was the 7th TI and he is nowhere near the top of the field. Miracle- for example would be considered one of the best players in the world right now, but there is also every other mid player that made it to group stages.

 

[Arms Longfellow]

^ It doesn't matter, the Dota 2 AI also beat Sumail who you could certainly argue is the best mid player in the world. I've accepted that the AI now has superiority over humans in that limited 1v1 matchup environment. It will be far harder for them to achieve their goal of creating a 5 man AI team that can beat the best team in the world in a regular game, where there are way more variables and creative decision making required.

 

[Abraxis]

...and with Firaxis' limited budget...

Does Firaxis have a limited budget? They develop one of the most successful, iconic, integral and longest running videogame series ever

 

[Japper007]

Have any ambitious programmers ever taken a crack at programming a Civ playing AI?

Yes they have; Arago, a German company, made an AI last year that can beat any human opponent in Freeciv (essentially civ2 and civ3 mashed together). They did it by using their "Hiro" desicion making AI, which used text based input from several human players to form one strategy by just trying everything suggested thousands of times until it could do it.

I was one of the players that submitted strategies actually. I don't know much about the actual inner workings of their program but I do know the process was extremely expensive and mostly a showcase for their problemsolving AI. Still the result was impressive, the eventual AI could execute naval invasions, bribe cities and run an economy tighter than even the best micro-ers in our tester group.

Their website for anyone interested: https://www.arago.co/ai/freeciv/

 

[mckinney156]

If the AI was smarter in Civ you wouldn't be able to beat it on the higher difficulty levels.

As for DOTA 1v1 doesn't really say much.