Donkey Kong Forum
General Donkey Kong Discussion => General Donkey Kong Discussion => Topic started by: homerwannabee on January 21, 2018, 08:37:32 am
-
Now, I'm not talking about programming a computer with specific moves to beat the World Record. What I'm talking about is a computer learning the game from scratch with only the idea of getting the highest score possible in it's program. I've seen talk about how these A.I.'s are beating people in fighting games, and games like LoL, but what people fail to take into account with these type of games is that speed matters in all of these games. If you are able to generate moves 3 or 4 times faster than your opponent, then of course you'll do really really well.
With a game like Donkey Kong though, speed means absolutely nothing. You actually have to learn the game, but learning the game is not enough. You need to get to the killscreen, and then relearn to score higher, and then relearn again to score even higher.
Also, another problem, The major programmers actually caring enough to put A.I. to the test of Donkey Kong. The major A.I. may not be programmed for that. It seems rather random the games they choose. So a lower level A.I. programmer may actually be the one who has the A.I. program to beat the Donkey Kong high score.
So based on everything I personally believe it's going to be around the year 2030 before an A.I. program actually beats the Donkey Kong Arcade/MAME world record.
-
Any significant effort to program an AI to break the DK WR would result in that record being CRUSHED in short order. Auto Machine Learning is pretty powerful and still in its infancy.
This could easily happen before the KO6.
! No longer available (http://www.youtube.com/watch?v=YNLC0wJSHxI#)
-
Then explain this video? 19 months ago this A.I. can't even get 15k on the Atari 2600 Galaxian game. The Atari 2600 game is WAY easier than the Arcade/MAME version. https://www.youtube.com/watch?v=AVg_YIp09ps (https://www.youtube.com/watch?v=AVg_YIp09ps)
-
Also this is current A.I. playing Super Mario Bros live. The computer has taken like 2 months, and it still hasn't completed level 1 of the game.
https://www.youtube.com/watch?v=QiXWqjNurQw (https://www.youtube.com/watch?v=QiXWqjNurQw)
-
That is FAR from state of the art AI, If google took on DK it would be solved by march.
-
That is FAR from state of the art AI, If google took on DK it would be solved by march.
Well that was another thing though I factored in. If the top level A.I. programmers ignored something like this, than the next lower level programmers would tackle it, and by the looks of things the lower level A.I. programmers may take till 2030.
-
the lower level A.I. programmers may take till 2030.
Maybe longer. Shopping exclusively at Whole Foods and telling you how much the music and beer you like, actually sucks, takes up a lot of time.
-
ROFL
-
I don't see an AI as they are currently and without human coaching coming up with point pressing moves such as jumping near DK's foot on the girders anytime soon.
-
That is FAR from state of the art AI, If google took on DK it would be solved by march.
Yeah if Alpha Zero was thrown at this it would likely be pretty quick all things considered. Probably a 1.3 million score within 3-6 months if that is what people wanted. Otherwise a KS would happen fairly quickly.
https://gaming.youtube.com/watch?v=QiXWqjNurQw
-
The interesting thing about AI is that it is potentially limitless. While AI might be just getting started you really have to look at the concept in the long term. What is AI going to be like in 100 years? What will AI be like in 1,000 years? In 100,000 years?
These time spans might seems to be unreasonable to understand, but the fact is, the Earth will be here 100,000 years from now and who knows what the heck its going to be like to live on this future Earth.
I have always wondered what future AI will be like. Will it just be one big super computer which knows everything? Or will there be individual humanoid robots that are equipped with AI and each robot has their own unique personality that is shaped by their experiences much like us humans?
A future humanoid with AI could beat the Donkey Kong world record, but then again you would have to ask the robot "why do you want to beat the DK record?". A future robot with AI would have to have the desire to choose to beat the DK WR. It would have to be a personal choice made by the robot. We could program and tell a primitive AI bot to set the DK WR, but I don't think that we could tell an advanced AI what to do period.
If an advanced AI were to choose to try and beat the DK WR. I would imagine that it would go something like this:
1) Downloads MAME and Donkey Kong rom off the internet
2) Plays the Donkey Kong game on MAME with the game sped up by 1,000 times or more inside its own operating system
3) Plays the game several billions of times with in a short time span
4) Sits down at Donkey Kong arcade machine and basically preforms a "perfect" Donkey Kong game first try
Seriously that's what it's going to be like. It will be unreal.
-
If AI were truly "learning" without human coaching, would it not be just as likely to find game breaking glitches to manipulate score, including any possible avenues for arbitrary code execution, as it would to maximize "normal" gameplay?
I imagine watching one of these AI performed games live and then we get to the kill screen and... what the...??? He's not dying?! Because somewhere along the way, without anyone even noticing, some tricky movements were done earlier in the game to rewrite how the level counter is saved in memory and BAZOOM, kill screen averted!
This would of course depend on if such an exploit exists in the DK code, I'm just imagining how something like this could play out in the case that it did- and in the case as I'm sure it will with other games. Imagine AI that finds major glitches for speedruns because, taking the heuristics of human cognition out of the picture, the AI doesn't "know" what normal play looks like and doesn't automatically make any "assumptions" on how game mechanics work.
This, incidentally, is probably why it's taking so long for AI to do tasks that we consider simple. Without a heuristical base for decision making, it has to account for significantly more possibilities than a human ever would- including things that would seem ridiculous, like backtracking to another screen just to deliberately fall into a hole, hitting a specific pixel at a specific time on the game clock in the process, and then dying. You'd never know if that would cause a wrong warp because you'd never try it. However, AI might be able to find it by brute force.
-
After rethinking the question, I came back here to clarify and correct some of what I said- only to find that KongTower already beat me to the punch! You are, of course, correct that I was missing the key component to what AI actually is, and that would be the I, or, intelligence. I was conceptualizing the technology in the same manner as a brute force method of solving a board or logic game- using software to run simulations of every possibility, discarding failed results and keeping sucessful results, to arrive at a clean decision tree that leads to perfect play. I was imagining, basically, a program that writes its own TAS for the game, which is not at all what we're talking about here, even if that would be interesting in itself.
Maybe a better analogy for AI gaming, and maybe KongTower can help verify or correct me here, would be if Johnny 5 took to playing Donkey Kong and kept playing and playing until he became better than all humans. The difference being that instead of a physical robot, Johny 5's play is generated by code that directly inputs controls into the game instead of using a joystick and buttons.
-
<confused>
-
It's actually more like how they created a girl in Weird Science than the traditional concept of what most people trained in AI from 1980's through mid-2000's had learned. It's only a couple years that this has been known as the breakthrough.
Wait....we can do this?
-
<confused>
^
-
But today, I'd love to hear a roll call from anyone willing to having had practiced harder stages in Donkey Kong by using MAME as a tool to build up motor-memory for the game.
Literally 99% of active members of this forum.
-
I'm not saying this to sound high and mighty, but to the best of my knowledge I am the only active player who has never played DK in mame. I always assumed that I was in a boat with only <Wiebe> and <Billy> in this regard (and maybe <Kuh> ), and that is the reason I chose the "long road" so to speak. It wouldn't surprise me to learn that all 3 of them have played on mame (specifically to hone their skills on certain screens), I'm sure there were plenty of 3-500K players back in the 80's and early 90's that took this route, if only by necessity.
To answer your question, If you counted <Wiebe> and <Billy> in the percentages, it would still be over 99% as Wes stated. Learning without mame is dumb and takes infinitely longer, from my experience. <YSG>
-
I've heard the argument that one shouldn't use emulator functions to practice and, similarly, I've heard that practicing Connect Four in perfect-play software is "cheating". (if the latter sounds out of place, allow me to add that C4 has been a fixture in my life and, when I'm de-rusted, am an expert player in the game). If it's simply a preference, like with f_symbols, that's one's prerogative.
To actually insist or demand that others do the same is just ROFL To me, it's like saying you can only practice kicking field goals during a real game or that you can only practice running during a race.
Thankfully, at least 99% of people here would agree with me and the whole argument is moot <Allen><Sanders>