calling all programers!!!!!!! well ne1 w/a good idea!

in Mac Software edited January 2014
Before i begin i know what i am doing is kinda ludacris but me and my best friend want to do it. In short we are making a starship game similar to myth. there connection is that you are given power over a small group of ships sent from mission to mission depending on the world you are given dif. ships, though your race still remains the same.

part one: the idea is not only do you control and tell them what to do, you can train subordinates. the subs can control a group of fighters while you look at the rest of the fleet. once a mission is accomplished then you can re direct them or assume control etc.

part 2: this is where you come in. we are going to use a really complex ai. actualy, a combination of ai's that interact together. responding as you and your subordinates would and learning what you do and adapting. the learning part and and adapting is why we are developing the game. we are going to use a combination of neural networks as wll as an algorithm following sun tzu on the art of war. ok here we go

the general flow of things goes like this, a battle is played naturally the human is expected to win. a replay of every battle gets saved to an output file that could pretty much be described as a textbook. each ship's, damage incurred as well as the positions and speed are put into the book. however, when the battle is over the first nn goes to work.

i know that neural networks take alot of practice, as well as cpu time, so we found a way to start working on this. the first nn called joshua is modeled after a type of netwoork that was tought to learn what a chair is, it worked by being told this is a chair, this pic is not a chair. after several tries it final was able to tell you a chair, be it a rocking chair or an office chair.

joshua works the same way, it looks at the batle log over and over, presented with the same information, but the question arises how do you teach when its doing well and when it geting its butt kicked. i was thinking of putting in simple variable that compared the rate of damage to it ships vs the human ships over dif periods of time and then analyzing weather its winning or losing. for example in time period 10 seconds to 30seconds humans lost 3 ships, ai lost 1, then startegy is working save position paths to warfile. if losing save position paths to warfile as losing.

to actualy execute, i was thinking of a style similar to the seti at home project. not that it would be distributed but that joshua would work in the backround with available cycles. so you could play a game come back the next day and it would be smarter.

next time when the game is loaded a new nn is loaded, named colossus. colossus is made from a dif pattern recognition nn. my friend did one last year that defines shapes either circle triangle or rectangle. the idea would be to use the warfile created by joshua to recognize when the human player is doing something similar to what he has done before. so if you sent two fighters north and 5 south and it worked, then it would see it as a threat if you sent 5 south and 2 north.

then comes ender, after the behavior is analyzed what is going on is sent to ender and ender takes over briefly. ender then loads tactics from sun tzu to control the actual ships. this is a part where i am still very shaky and need some more help.

oh yeah i had been thinking it would be kewl if you could trade profiles on the net and let other people play ur computer that you trained or etc. or even combine the files to build like an uber ai. oh well its late.

now that you know what im trying to do, if there is any help suggestions or any people i could talk to then i t would be most appreciated, and please links are fine but people are better. oh yeah sorry for the long post but i this is a weird subject.

[ 01-24-2003: Message edited by: myahmac ]</p>


  • Reply 1 of 8
    der kopfder kopf Posts: 2,275member
    I was involved (not deeply, just a bit) in the Neural Networks department at the University Ghent, Belgium, last year as a comp. sci. student. They have a neural networks computer, the CAM-Brain, available there, and I actually made that available (somewhat, at least) through the internet. I haven't exactly studied neural networks that deeply, but I bet the fellow I worked with, a PhD student (working with neural networks), wouldn't mind looking over your stuff. And if he would mind, you'd still not have lost anything asking. If you're interested, check out <a href=""; target="_blank">this page</a>. The guy is called Hendrik Eeckhout and his email is in the contact section.
  • Reply 2 of 8
    scottscott Posts: 7,431member
    I'm not sure you can use an NN in that way. My experience with NN has been with them being used as classifiers, eg is this blob in a 64x64 sub set of the this image cancer or not. But blob is not just "blob" but a set of quantified characteristics of the blob.
  • Reply 3 of 8
    myahmacmyahmac Posts: 222member
    the idea is pattern off of a chair. a chair has such a ixel at 1.1 and another at 1.2 but a dif at 1.3. then it is told weather it is flase or tru. how ever the program not the nn decides weather the position increment are apattern or not.
  • Reply 4 of 8
    scottscott Posts: 7,431member
    I have heard of NN that use "all" pixels as inputs. I don't think they are that good and require a lot of training data. That brings up another issue? What are you going to use as training data?

    The chair example is a classification issue. Is this a chair or not. More to the point, score this set of inputs to estimate the level of chairness on a scale from 0 to 1.0. Then you use ROC analysis determine how good the NN is.

    But you know my knowledge of NN is limited to those type of problems. So ... I may be myopic on this issue.
  • Reply 5 of 8
    well as traing data i was thinking of fighting with the ai turned off, other than the firecontrol. for example several atack head on vs flanking by coming to the left. and then flanking by coming to the right etc. stuff like that should get it used to 3d space right? then for repetitiveness, could not the nn study the same data and be given the same result as well. part of the game ai has to know when its not doing well, so i could use that factor as a judge of eithe ryes, it works, or no it doesn't.

    [ 01-24-2003: Message edited by: myahmac ]</p>
  • Reply 6 of 8
    maskermasker Posts: 451member
    Sounds like acool project that is beyong my mental abilites.

    i'm assuming that there is some way to sort out the fact that the losing strategy used today may be tomorrow's winner?

    Also, (being a long time myth player) it seems that success is very dependent on allocation of units and timing and reinforcements. Can you tech the NN to learn when do allocate reinforcments to a battle to swing it to victory?

    There a re also distractionary battles that pul units away from a disguised master battle before you strike.

    just how smart are you wanting this AI to be?

  • Reply 7 of 8
    myahmacmyahmac Posts: 222member
    what you mention are actualy very good points masker and go into tactics that the ai would need. we are basing alot of those built in off of sun ztu the art of war. as you say distraction is key and then isolation etc. another point we ar bringing in is orientation. the setting is in space where down is where you want it to be. most people will set up down just like in the room they are playig in. but to the computer, the enemy is always down. so that line abreast will look diferent and may conceal units if the layer does not also learn to adapt to the camera. the idea is we want the ai to learn as well as the human player. and yes we want it to be verysmart. any suggestion like i said is helpful, as of now we ar setting up like in the matrix "a construct" to load are obejects in such as ships and etc. then we will begin testing basic parts of the ai and expanding the user interface. but we can always use more strategy tips. in fact give us what you would want in a strategy game.
  • Reply 8 of 8
    scottscott Posts: 7,431member
    I would suggest that you look into the AI that used to program chess games. I'm still not seeing how ANNs can be used in this way.

    [ 01-25-2003: Message edited by: Scott ]</p>
Sign In or Register to comment.