LinkBack URL
About LinkBacksRead more.Multiple AI agents learned CTF mode from scratch and leveraged collective intelligence.
Read more.Multiple AI agents learned CTF mode from scratch and leveraged collective intelligence.
Good, put this into robotics so that next time war breaks out we can send those onto the battlefield instead of a bunch of innocent teenagers
....And when they begin killing innocent, now defenseless teenagers?Originally Posted by EvilCycle
Good, put this into robotics so that next time war breaks out we can send those onto the battlefield instead of a bunch of innocent teenagers
as someone who still plays Q3 can they release this as some less predictable single player would be ace.
Teaching AI how to kill and capture...
Games have unbreakable rules. Combat does not. 'Nuff said...
The AIs are certainly good but I have to wonder how excellently the best human players would play if they were able to gain 150,000 to 450,000 games worth of experience! 150K is only 40 games per day, every day of the year for ten years. Any volunteers? ;-)
In games like Go and Backgammon, the nets are given a game state (basically the board and pieces) and trained to give as accurate an evaluation of that state as they can. Given a state and selection of possible actions, the resulting state of each action is determined and thus its evaluation can be obtained (estimated by the net). The action that gets made is the one with the highest evaluation. This is what AIs that play Go and Backgammon do. They have no understanding of their respective games.
They don't know the concepts of, for example, hitting a piece in Backgammon, or creating a certain configuration to acquire territory in Go. All they know is that when they've encountered states which represent the outcomes of those tactics (without knowing that they *are* such outcomes), the evaluation is X .... and X happens, after sufficient training, to end up higher than the outcomes of other states that can be reached from common predecessor states. Hence the AI will make the hit or strive to achieve the territory without needing to know what it's doing or why. Indeed, such an AI cannot explain its conceptual logic, for there is none; it can only say "I did that because, in looking at all the possibilities, that one had the highest evaluation".
The game-learning concept is the same with video-game playing AIs so it's no surprise that they are showing similar capabilities. The games states are more complex - a *lot* more complex - because instead of a board with simple possibilities (no piece, my piece, your piece), you have a grid of pixels which represents a visual view upon some part of an underlying model (terrain height, object types and locations). This grid of pixels has to be translated into an equivalent model state. But from that point on, the exercise of learning to play and win is no different than with a board game - associating states with increasingly accurate evaluations.
It's certainly an achievement to have developed a perceptual pipeline that can convert pixels into a model state. It's also very impressive that, unlike a board game, where you only have a few options such as "Place piece X at location P" and each player makes just one action per turn, video games have much wider options and multiple actions can be made in sequence, for example "select weapon W, run to location P, fire at location (x, y, z) then crouch" .... but while it's an "AI" doing this, it's hard to credit it with intelligence.
A human player knows when to select a BFG because it's a *BF* gun which makes *BF* holes and that's what's needed to get past the giant troll just ahead. The current AIs know that in a game state where the bits that arbitrarily encode the presence of (what a person would recognise as) a troll, and where among the action options there's one which selects (what a person would be able to recognise as) a BFG, the outcome which selects that weapon has a higher-valued outcome than actions which don't select it. And they discover this by trying *all* the actions possible from this point and looking up the evaluations, not by knowing anything about BFGs and trolls.
The best Backgammon bots in the world "explain" and "teach" the best way to play the game by giving a list of possible moves along with the corresponding evaulations. That's it. No explanation of anything backgammony is currently possible so the greatest benefit goes to players who are comfortable with numbers.
*True* intelligence, in my opinion, will only begin to emerge when a game-playing AI knows *conceptually* what it is looking at, what the various actions do in their own right, why to use them in certain circumstances, what the concepts behind the various goals of the game are, and even more convincingly, when it can explain its actions in those terms.
I've not read their research paper but doesn't the impressiveness of beating human players depend on how good those humans are? If we're talking about a selection of the worlds top 100 competitive Q3 players then colour me impressed, if on the other hand we're talking about the average player found on public servers then it's not so impressive.
If the AI was a camping sniper than it would easily thrash anyone!!
Last edited by CAT-THE-FIFTH; 05-07-2018 at 06:31 PM.
So basically, the AI has an equation which lists moves and so on. The AI tries these and moves performed at certain times or in certain patterns have an outcome closer to the desired. Therefore the AI weights the equations so it performs moves like this more often at such and such a time. It gradually improves this weighting over time and thousands of games. This is a massively complex equation which has to be created but the concept when broken down seems very, very simple. It shows no understanding of what's actually happening but it's just modifying weighting in an algebraic form to get closer to a desired outcome. Zero comprehension, just maths with an answer and the shuffling of variables to make the output match a desired state.
Fundamentally, am I wrong? I have no understanding of AI and have done no work in the field but this seems pretty simple.
^^That's^^ basically sums up what grinds my gears with 'AI'. What everyone is calling 'AI' is not much more than a collection of pattern matching algorithms and weighting, just maths with an answer and the shuffling of variables to make the output match a desired state.
Maybe I'm being overly harsh but to me intelligence is the ability to acquire and apply knowledge and skills to something you've never done before, we can teach a pigeon to peck at a flashing light but IMO that doesn't make it intelligent, it just means that through trial and error its learnt that pecking the light means food.
That is general intelligence, and certainly no-one knows how to make computers do that (yet).
What we have here is something more like "flying by the seat of your pants" muscle memory. It is useful in itself, and the techniques involved will no doubt evolve at some point into creating the feedback for a general intelligence network.
So yes, I think you are being a bit harsh. If all the AI out there stopped working, I think you would be shocked at how much we depend on it. For starters, you wouldn't be able to Google what was going on...
Don't get me wrong, I'm not saying these programs are not useful and probably very sophisticated, i agree with your sentiment that we depend on them, what i take exception at is calling it artificial intelligence when in my mind it's nothing of the sorts, as i said we can train a pigeon to do simple tasks but i think it's safe to say we wouldn't consider that intelligence, but when a computer (with the aid of programers) puts together what are in essence lots of pigeons pecking things we call it intelligent.
It used to be called machine learning, I can only assume marketing got involved and thought AI sounded sexier.
Corky34 (06-07-2018)
On the topic of AI:
Blue Yonder, 06.07.18 – Retail artificial intelligence (AI) specialist Blue Yonder has turned the attention of its powerful AI solution to predict the results for the remaining games of the World Cup 2018. The AI powered algorithm, developed using techniques honed at CERN, puts England on course for a win on Saturday’s match against Sweden, though only gives the team a 9.8% chance of taking home the ultimate prize.
Commenting on the predictions, Christian Haag, Data Science Consultant at Blue Yonder, said: To arrive at the predictions, Blue Yonder’s powerful AI solution analysed every international football match played since 1872 (approximately 38,000 matches) running over 1 million simulations of the World Cup. From this, Blue Yonder has identified Brazil as the overall winner, though it is set to be a tight race between the remaining countries.
Blue Yonder’s latest predictions for the winner of the Fifa World Cup 2018 are as follows:
Brazil – 29.8%
France – 15.3%
Russia – 11.5%
Belgium – 10.1%
England – 9.8%
Uruguay – 8.8%
Croatia – 7.6%
Sweden – 7.1%
I've only seen the name AI used for machine learning with neural nets and generally with deep learning, which are currently very trendy. There are lots of other machine learning techniques out there, so calling it ML would be overly broad. I can't see you ever getting a Bayesian spam filter to play Quake.
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote