1002: "Game AIs"

This forum is for the individual discussion thread that goes with each new comic.

Moderators: Moderators General, Prelates, Magistrates

mfb
Posts: 950
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: 1002: "Game AIs"

Postby mfb » Wed Mar 29, 2017 10:30 pm UTC

I was asking about SC, and the SC AI duodecimus mentioned in particular. I know that computing power in general can be an issue, but it does not have to.

duodecimus
Posts: 43
Joined: Tue May 19, 2015 12:25 pm UTC

Re: 1002: "Game AIs"

Postby duodecimus » Mon Jun 19, 2017 6:24 pm UTC

It isn't an issue in starcraft ais. The game is pretty small.

So much of a non issue that we can speed up the game by 50-100 times speed. We were really only limited by how fast we could render frames.

Some of my anti-stealth work and a bad ''sell all buildings before they get destroyed'' script would slow it down, but the graphics were by far the limiting factor in gamespeed. I think we had a 20 ms timeout for bots getting disqualified.

If a game went more than a few hours, which was really rare, the build order decision maker would start to choke because it was something like n^4 on a list of all past and current units.


It didn't happen much in our class, but the national contenders would change strategy based on the early builds of the opponent, and would remember what a bot last played, and what they last used, and etc. All the stuff you can hear about rock-paper-scissor AI competitions stays valid here.
It was largely just 'build X to counter Y' though. Stealth units were pretty popular as a feint actually, since they're expensive to counter.

User avatar
Flumble
Yes Man
Posts: 2249
Joined: Sun Aug 05, 2012 9:35 pm UTC

Re: 1002: "Game AIs"

Postby Flumble » Sun Jan 27, 2019 2:55 am UTC

Good news everyone!
AI has effectively beaten StarCraft 2. Here's a video of the matches and stuff and here's the blag.

It took a bunch of people a bunch of months with a bunch of state-of-the-art techniques on a bunch of really expensive computers, but it beat two pretty good gamers 5-0. And then lost a match after learning with a new restriction.

mfb
Posts: 950
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: 1002: "Game AIs"

Postby mfb » Mon Feb 04, 2019 9:48 am UTC

The AI plays unconventional - humans often copy strategies from other humans, but this AI is more creative. It was mainly trained by playing against itself.

As far as I understand they trained the AI on a single map and only protoss vs. protoss, but I guess the same AI would work well with other maps and races, too, if trained for them.

User avatar
Soupspoon
You have done something you shouldn't. Or are about to.
Posts: 4060
Joined: Thu Jan 28, 2016 7:00 pm UTC
Location: 53-1

Re: 1002: "Game AIs"

Postby Soupspoon » Mon Feb 04, 2019 12:21 pm UTC

Unless there's a peculiarity of that setup that lends itself to the AI meta-plan, I suspect that's the point. It optimises to the specifics (or generalities) it is trained to. If you change the specifics (and generalities) to somewhere else in the same general sphere and re/super-train it then it should be good at the new (or, possibly, additional option of) outlook.

Specificity in relation to the problem counts. An AI good at chess may not be any good at draughts(/checkers) if it is keyed specifically towards chess. I don't see that as a problem where "same general mechanics, different strengths" play happens, and maybe the chess machine can be grounded enough to 'relearn' the draughts-piece movements (jumping mini-bishops?) if given that latitude. It could even be capable of learning both boards, though would have to further learn that one such game never morphs or mixes with the other, even while it holds a Net of understanding for both (a pawn reaching the other end never turns into a backwards-capable hopping mini-bishop!).

Recognising the board in front of it and focussing on what it has learnt of that board (chess, draughts, go, backgammon, hungry-hungry-hippos, etc) is part of the skill that a truly general AI must be capable of. Recognising which race-type it is controlling (and race(s) competed against?) is a lesser but still necessary twerk if you want to overlay multiple areas of learning on the same back-end, and the argument now is whether the complexity of stored knowledge merely multiplies by a small fraction of the different number of 'distinct playing combinations' (given the sharing) or if the whole "hang on, if you're considering RACIAL SPECIAL MOVE then don't - you don't have that if you're OTHER RACE" boundary logic in the knowledge-base plumps up the 'experience' memory (however it retains its learnt strategies) and needs a lot more training to fully establish that 'logic' (like how a chess program might need to learn that castling is not possible beyond certain bounds, if it is not pre-built with that rule, which is a whole order of magnitude more difficult than supplying it with the "valid move guidance" from the start, but shows a greater ambition for the AI than merely doingma variation on probability trees).

User avatar
ucim
Posts: 6860
Joined: Fri Sep 28, 2012 3:23 pm UTC
Location: The One True Thread

Re: 1002: "Game AIs"

Postby ucim » Mon Feb 04, 2019 3:32 pm UTC

How is recognizing what game it's playing conceptually different from recognizing what piece it's moving?

Jose
Order of the Sillies, Honoris Causam - bestowed by charlie_grumbles on NP 859 * OTTscar winner: Wordsmith - bestowed by yappobiscuts and the OTT on NP 1832 * Ecclesiastical Calendar of the Order of the Holy Contradiction * Heartfelt thanks from addams and from me - you really made a difference.

User avatar
Soupspoon
You have done something you shouldn't. Or are about to.
Posts: 4060
Joined: Thu Jan 28, 2016 7:00 pm UTC
Location: 53-1

Re: 1002: "Game AIs"

Postby Soupspoon » Mon Feb 04, 2019 4:05 pm UTC

For the sufficiently generalised AI, not at all. To the extent that an AI with a camera can use trained image recongition to work out if it's sitting down at a game of tiddleywinks or poker, and which type of poker game (or at least subset).

But if your human handlers are skipping various layers of incidental interfacing issues and is supposed to just give you a simplified current board state array having already ensured it's "the right game", and not wasted time having you deduce the necessary conditions of en passant from a review of a million 'training' games or an "invalid move!" external feedback module that is essentially the same thing, then you're probably going to be a specific-aim AI that can't handle even Two Queen Chess so easily.

elasto
Posts: 3757
Joined: Mon May 10, 2010 1:53 am UTC

Re: 1002: "Game AIs"

Postby elasto » Mon Feb 04, 2019 7:25 pm UTC

How do humans learn to play a random game? It's via feedback/instruction from others who have already played it, typically ahead of time, and sometimes via a practice game.

Mere hours after being told the rules of chess and nothing else, AlphaZero crushed a chess AI that had been refined over the course of years, an AI which itself can beat the best humans.

It seems plausible that AlphaZero isn't far off being able to self-learn anything a human can, given no more instruction than a human would need (and probably much less, most of the time).

rmsgrey
Posts: 3632
Joined: Wed Nov 16, 2011 6:35 pm UTC

Re: 1002: "Game AIs"

Postby rmsgrey » Mon Feb 04, 2019 7:59 pm UTC

elasto wrote:How do humans learn to play a random game? It's via feedback/instruction from others who have already played it, typically ahead of time, and sometimes via a practice game.

Mere hours after being told the rules of chess and nothing else, AlphaZero crushed a chess AI that had been refined over the course of years, an AI which itself can beat the best humans.

It seems plausible that AlphaZero isn't far off being able to self-learn anything a human can, given no more instruction than a human would need (and probably much less, most of the time).


Chess is a game that plays toward traditional AI strengths rather than human strengths. Show me an AI that's good at Codenames or Pictionary, and I'll be a lot more impressed than by one that's good at Go.

elasto
Posts: 3757
Joined: Mon May 10, 2010 1:53 am UTC

Re: 1002: "Game AIs"

Postby elasto » Mon Feb 04, 2019 11:57 pm UTC

Chess is a game that plays toward traditional AI strengths rather than human strengths. Show me an AI that's good at Codenames or Pictionary, and I'll be a lot more impressed than by one that's good at Go.

Sure, but I'm old enough to have heard that argument many times over.

AI gets good at X, people say 'Pfft. X is easy; Get back to me when it can do Y'. Then when it does Y, someone else says 'Pfft. Obviously AI can do Y. Z is where things get really interesting'.

It's as if AI is seen as dumb, therefore if it manages to do something, the only explanation must be that the thing it did actually wasn't that hard after all. When an AI can win at Pictionary, there'll surely be someone posting 'Pfft. That's nothing. Get back to me when an AI is self-aware and enslaves the human race!'

Image search, real-time language translation, self-driving cars... all these things would have blown the socks off a researcher of even two decades past; But AlphaZero is a gamechanger - not because of what it conquered but how it conquered it: By teaching itself.

We really could be on the brink of an AI that can design an improved version of itself, at which point all bets really are off...

At the Google I/O developer conference this month, there was one super clear message: Alphabet is now in the artificial intelligence business. This has big implications for lots of people.

"We spoke last year about this important shift in computing, from a mobile-first to an AI-first world," said Google CEO Sundar Pichai in the opening keynote, setting the stage for how Google is adding AI to everything.

"We are excited about designing better machine learning models. What better way to do this than getting neural nets to design better neural nets? Whenever I spend time with a team and think about neural nets building neural nets, it reminds me of one of my favorite movies, Inception. I tell them, we must go deeper--across a wide range of disciplines."

rmsgrey
Posts: 3632
Joined: Wed Nov 16, 2011 6:35 pm UTC

Re: 1002: "Game AIs"

Postby rmsgrey » Tue Feb 05, 2019 5:06 pm UTC

elasto wrote:
Chess is a game that plays toward traditional AI strengths rather than human strengths. Show me an AI that's good at Codenames or Pictionary, and I'll be a lot more impressed than by one that's good at Go.

Sure, but I'm old enough to have heard that argument many times over.

AI gets good at X, people say 'Pfft. X is easy; Get back to me when it can do Y'. Then when it does Y, someone else says 'Pfft. Obviously AI can do Y. Z is where things get really interesting'.

It's as if AI is seen as dumb, therefore if it manages to do something, the only explanation must be that the thing it did actually wasn't that hard after all. When an AI can win at Pictionary, there'll surely be someone posting 'Pfft. That's nothing. Get back to me when an AI is self-aware and enslaves the human race!'


Or, alternatively, it's as if different people have different thresholds for what it takes to be impressed by AI. Some people were probably impressed by a computer able to play Pong, others by Noughts and Crosses, others by Chess and Go, others are holding out for something capable of consistently passing the Turing Test, others wouldn't be impressed by an AI that successfully enslaved the human race (it's not possible to prove self-awareness beyond all doubt).

There is also an element of AI research historically having solved problems that were thought of as "hard" by finding simple solutions. One of the things that's come out of AI research is a realisation that a lot of things humans think of as "hard" are actually much simpler than we realise, but a lot of the things we think of as "easy" (or don't think of at all) are actually much more complicated.

There's also a difference between a computer being given the rules of Chess and told to learn the game, using a procedure for learning how to play well that it's been given; and a computer coming across the rules of Chess and deciding to learn the game and figuring out how to come up with good heuristics.

User avatar
Pfhorrest
Posts: 5447
Joined: Fri Oct 30, 2009 6:11 am UTC
Contact:

Re: 1002: "Game AIs"

Postby Pfhorrest » Wed Feb 06, 2019 3:43 am UTC

Did someone say Pictionary?
Forrest Cameranesi, Geek of All Trades
"I am Sam. Sam I am. I do not like trolls, flames, or spam."
The Codex Quaerendae (my philosophy) - The Chronicles of Quelouva (my fiction)


Return to “Individual XKCD Comic Threads”

Who is online

Users browsing this forum: Google Feedfetcher and 39 guests