Its Just Too Massive!

Discussion in 'Planetary Annihilation General Discussion' started by stevenrs11, January 17, 2014.

  1. keterei

    keterei Active Member

    Messages:
    258
    Likes Received:
    93
    Let me just say, that there is a difference between loose and lose. Loose = the opposite of held together. Lose = to have something taken or lost. I think by 'mass points' you mean metal nodes? It sounds like you're asking that PA become an easier game because you don't like how much it demands your attention. We're fighting on planets, how can you not expect to have your attention stressed? That is what makes PA what it is, and that's how I (and safe to say, others) like it.
    Last edited: March 7, 2014
  2. stevenrs11

    stevenrs11 Active Member

    Messages:
    240
    Likes Received:
    218
    I don't expect PA to become easier, or demand any less of my attention. All I wish is that I could spend more of my attention doing things that aren't either building more mass points, or spamming rows upon rows of factories to spend it all. Right now, doing that well WINS, and dumping all of your attention into that is the best way to play.

    I think you can do that without slowing down the game, or changing into something else. If anything, it would make the game harder. Expanding is easy to decide to do- When? Always. How much? There cannot be too much. There are no important decisions when it comes to expansion. It just requires a disproportionally large amount of attention to actually do it.

    If you want actual strategy, the decisions themselves need to be hard. The judgment calls, like how to assault a fortified position, or which type of unit to focus on, or how to use terrain to your advantage- these things are hard decisions to make that are easy to actually go and do.

    If, going into the game, you already have made all the important decisions, then the game simply becomes 'who can execute faster', and that isn't all that fun.
  3. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    Like stevenrs11 said, even recognizing the state of the game in PA is very hard. Even if the computer have memorized all the gambits in Chess or what to do with the current board, you can simulate each move, calculate the strength of the board positions and from there branch out and simulate the next move, calculate the strength board position and so on which will take you quite far in chess. This could be seen as brute forcing. You simulate all the possible moves and calculate favourable moves in that regard.
    However, that isn't possible in PA because if you can view the number of possible moves to reach infinity from just one second to the next second.
    How many possible "move"-orders can you give to a Dox?
    Is that number limited or infinite?
    In the next frame you can change the "move"-order to that Dox.
    How many possible combination of moves are there in a single second for a single Dox? Infinite?

    I'd say that weighing the probabilities in PA is harder than in poker.

    I am convinced that computers will be able to learn and play RTS better than humans one day.
    However a computer isn't smarter than the software that runs on it so until we have some major AI-breakthroughs I don't think it is feasible to expect AIs to be better than humans in largescale simulation based RTSes.
  4. liltbrockie

    liltbrockie Active Member

    Messages:
    314
    Likes Received:
    160
    Pretty sure its limited.
  5. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    Limited to what positions? Does the discrete amount of positions depend on the size of planet as well?
  6. Dementiurge

    Dementiurge Post Master General

    Messages:
    1,094
    Likes Received:
    693
    Agreed.
    The software is only as good as the knowledge of the programmers who made it. My personal theory is that it's really our understanding of how games are won and how beings* learn that limits the ability of AI to grow.
    The reason AIs have to brute force Chess is because, at a conscious level, we only know how to brute force Chess... And at a sub-conscious level, no Grand Master could explain to you what's happening.
    Last edited: March 7, 2014
  7. liltbrockie

    liltbrockie Active Member

    Messages:
    314
    Likes Received:
    160
    Oh yes very much so
  8. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    That's quite a big amount of possible moves and then when you take all the combinations of 2 different move orders... Anyway. Where are you going with this?
  9. cdrkf

    cdrkf Post Master General

    Messages:
    5,721
    Likes Received:
    4,793
    I can see Sorian surprising you... The way the ai is going to function in pa is very different to the chess analogy for the reasons you've stated. The pa ai is being developed around a neural network that will allow the ai to efficiently assess a situation and even adapt to balance changes.

    My personal bet is that when given a resource advantage (i.e. hard) the ai will be capable of beating experienced, competent players. Remember the ai already has access to more apm than a normal player. Sorian has now added some good unit micro as well...
  10. liltbrockie

    liltbrockie Active Member

    Messages:
    314
    Likes Received:
    160
    No where. I'm not even sure what this thread is about lol
  11. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    I am aware that Sorian is using neural networks for his AI.
    I think a good analogy is that the AI exercises by playing against itself and do some random changes to its' gamestate analysis and behaviours where Sorian is the teacher and basically sets up the parameters of how the AI should evaluate the gamestate. It is very different from "brute-forcing" the game simulation.

    Well a resource advantage is cheating so it doesn't count. Any AI could win in PA if they start with 2 nukes over the opponents commander.
    Similarly it is also really easy to make an AI better than a human at a game when fast input is required. Making an AI better than a human when micromanagement isn't a big factor is much harder.
  12. cdrkf

    cdrkf Post Master General

    Messages:
    5,721
    Likes Received:
    4,793
    Well I think that the pa ai is going to be good and a worthy opponent once finished. Sorian knows his stuff.
  13. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    Yeah, I think so too but is it going to beat top players without cheating? I doubt it.
  14. cdrkf

    cdrkf Post Master General

    Messages:
    5,721
    Likes Received:
    4,793
    I guess it depends on what Sorian can do with the 'top' level management. The ai is going to be very strong with individual units and squads. What I'm hoping is that he can add some special sauce to get it reacting to the player as currently it attacks but it doesn't counter. For example if the ai scouts and spots t2 air it should adjust its build to counter it. I think it's possible for the ai to be good enough to tackle anyone 1 v 1, ffa games and so on I think will be harder due to the increased number of threats and so on.
  15. MrTBSC

    MrTBSC Post Master General

    Messages:
    4,857
    Likes Received:
    1,823

    actualy this one is a pretty good idea .... you can decide to just have a general buildorder AND issue low level buildorders that you consider higher priority ...
    f.e. you are building up general vehicleproduction and eco ...in that case you assign a general buildorder for all fabricators and the ai spreads the fabbers by what has priority by default like say first power then production then defense etc ..
    but now suddenly you see your enemy getting close in so you decide to take like 3 fabbers and order those to built a quick defenseline
    or you had that defenselineorder set up already but by ordering the fabbers you prioritise that order through assisting ... the way i imagine that is to allow players to assistclick the nonstarted blueprints of structures ... because you still would like to have the ability to prioritise structures without having to place them again
    but i imagine that to be difficult to implement .. this could also lead to problems in teammatches with shared army ...
    Last edited: March 7, 2014
    cdrkf likes this.
  16. arsene

    arsene Active Member

    Messages:
    166
    Likes Received:
    114
    You can brute force in PA just like you can in chess. Instead of trying to calculate gamestate you just exploit infinite apm. I'm confident that IBM could make a starcraft ai in two years that could beat all top players.
  17. godde

    godde Well-Known Member

    Messages:
    1,425
    Likes Received:
    499
    I would also like to control groups of units and individual units on a higher level.
    Kiting is a perfect example. A computer can manage many groups of Gil-E kiting simultaneously while it is hard for a human player to do the same. By automating kiting, humans can perform kiting much more easily and the need for micromanagement decreases.

    That hinges on that proficiency of micromanagement is important to the outcome of the game so that the computer can actually exploit infinite APM.
    In Starcraft that might be true but PA strives to reduce the importance of micromanagement so this type of "brute-forcing" might not work or might not make a big difference in PA.
  18. arsene

    arsene Active Member

    Messages:
    166
    Likes Received:
    114
    I'm curious, how much apm do you think top players should possess to be competitive? Do you think it should be sub-150? (i.e. accessible to anyone that practices a few hours per day) and where do you think the number is at now?

    Note that in starcraft it's around 300.
  19. stevenrs11

    stevenrs11 Active Member

    Messages:
    240
    Likes Received:
    218
    I uh, didn't say that, though I largely agree.

    As for brute forcing chess, its interesting if you read about how actual world championship level players play. Their basic strategy is very, very similar to how computers play- they memorize the largest possible set of positions and board states, and simply try to stay within those. If they go of into un-memorized territory, they often will try and force a draw.

    I definitely do not claim any authority on AI programming, but I have done a *very* little.
    At the most basic level, you have 'look ahead', and 'static analysis'. Look ahead determines how many steps forward into the sim you go, and static analysis determines the probability of winning based solely on the state at single time. From there, you build a tree that represents possible moves from a starting state. Now, its not all possible moves- you prune the tree based on lots of different factors, and realizing dead ends quickly is one of the single most important things in building an AI.

    For board games, it all breaks down very nicely. A turn is one step in the sim, and each piece has a very defined state. With an RTS like PA, you have to do lots of far harder things. You need to figure out how long a meaningful step actually is, how to represent the state, and then how to meaningfully get a static evaluation function for it. I cannot even imagine how complex this gets.

    Honestly, this model might not even apply to PA- I would love to have @sorian chime in a bit on how the AI works, though it might just explode my head. I imagine that micro of individual units would be separate to some degree from the overall eval- its something that groups of units would always do by default, and would be factored into the total tree as a sort of constant.

    Aaaand we are now very off topic. Oh well, its fun talking about AI stuff, and I bet that its easier to make an AI that expands well and spends money quickly than is actually clever. If expanding/spending mass remains the massive APM sink that it is currently, then I can easily see an AI doing that better than a person could ever hope.
  20. cdrkf

    cdrkf Post Master General

    Messages:
    5,721
    Likes Received:
    4,793
    In spring the rate can be very low for the top players. The spring ui is very modifiable and mature though. Also I wonder what definition people use for apm. Spring breaks it down to mouse clicks, keyboard presses and actual commands issued. I'd count the latter as most important in which case 30 is enough to play well, 60 for fast (not necessarily the best players) and I've never seen much over 100 and that was from a starcraft champ. I think the 300 your referring to is pure inputs I.e. clicks and key presses. A good ui can negate allot of that and pa is doing all the right things on that score.
    godde likes this.

Share This Page