i hope to be able to release a first version soon after the release of the next ta spring version (which comes with my extended ai interface)
2 preview pics of AAI
Moderators: hoijui, Moderators
2 preview pics of AAI
These are two pictures of the first working (= builds units and tries to attack enemy) version of my ai called AAI
i hope to be able to release a first version soon after the release of the next ta spring version (which comes with my extended ai interface)
i hope to be able to release a first version soon after the release of the next ta spring version (which comes with my extended ai interface)
Any specific features we should be looking forward too other than an AI with a wishlist?
And you changed the itnerface so untis knew what killed them, and mentioned something about using ti to get statistics?
I have a few experiments of my own and I'd be itnerested in seeing what you've done, otherwise continue with the hype. I shall release NTAI 0.27 shortly with that shot of the krogoths.
cookies for submarine
And you changed the itnerface so untis knew what killed them, and mentioned something about using ti to get statistics?
I have a few experiments of my own and I'd be itnerested in seeing what you've done, otherwise continue with the hype. I shall release NTAI 0.27 shortly with that shot of the krogoths.
cookies for submarine
AAI doesnt use any predefined buildscripts or hardcoded unit defs at all;
it reads everything it needs directly from the unit definitions and sorts units into groups (like mexes, power plants, scouts, etc.)
the only thing it has to know is the name of the starting units of every side; i tested it with xta, uberhack and aa - it worked - however it crashes with ta:wd
the ai keeps track of which units killed how many other units and that way calculates the efficiency of a unit - yet i dont know if it works because the ai has to play for a long time in order to figure out which units are useful or not...
it reads everything it needs directly from the unit definitions and sorts units into groups (like mexes, power plants, scouts, etc.)
the only thing it has to know is the name of the starting units of every side; i tested it with xta, uberhack and aa - it worked - however it crashes with ta:wd
the ai keeps track of which units killed how many other units and that way calculates the efficiency of a unit - yet i dont know if it works because the ai has to play for a long time in order to figure out which units are useful or not...
-
Lindir The Green
- Posts: 815
- Joined: 04 May 2005, 15:09
OMG, I think that IS what he's saying!!! WTF!!Kuroneko wrote:Wait, are you saying the AI can learn about the units?
Dude, PLEASE get this working soon! Will the AI save what it learns for later games? If so, we could have a real, learning, adapting AI for Spring!!!
Holy crap, imagine the possibilities...
PS: Please let your first release make expand /w water units... please...
An AI learnng from previeious games .... That sounds sweet I hope the AIs will start using water units and stuff soon Because I love Island battles 
.
So when u say the AIs will start learning from previous games do u mean something like it will learn how to attack players in new ways and attack in differnt positions, Set up traps for players like Setting up mines, Attack differntly every time you play that Same AI ... stuff like that?
it sounds like A lot of difficulty but if that can be done that would be outstanding!
.So when u say the AIs will start learning from previous games do u mean something like it will learn how to attack players in new ways and attack in differnt positions, Set up traps for players like Setting up mines, Attack differntly every time you play that Same AI ... stuff like that?
it sounds like A lot of difficulty but if that can be done that would be outstanding!
dont overestimate learning effect right now, i hope to be able to improve that in the future.
in the beginning, all units are considered equally effiecient against other unit categories; it saves its updated values after each game, so after some time, pewees or aks get much lower eff. values - stumpies or raiders higher ones
(i´m testing only with lvl 1 units right now..)
the drawback is that the ai has no idea which units are strong after deletion of its learning files, the whole process then starts all over again
next step will be to implement learning about the different maps (start locations, imporatant locations, etc)
in the beginning, all units are considered equally effiecient against other unit categories; it saves its updated values after each game, so after some time, pewees or aks get much lower eff. values - stumpies or raiders higher ones
(i´m testing only with lvl 1 units right now..)
the drawback is that the ai has no idea which units are strong after deletion of its learning files, the whole process then starts all over again
next step will be to implement learning about the different maps (start locations, imporatant locations, etc)
submarine wrote:dont overestimate learning effect right now, i hope to be able to improve that in the future.
in the beginning, all units are considered equally effiecient against other unit categories; it saves its updated values after each game, so after some time, pewees or aks get much lower eff. values - stumpies or raiders higher ones
(i´m testing only with lvl 1 units right now..)
the drawback is that the ai has no idea which units are strong after deletion of its learning files, the whole process then starts all over again
next step will be to implement learning about the different maps (start locations, imporatant locations, etc)
I want your child.
Seriously tho, noone ever bothers to make AIs that learn anymore. This is a huge step in the right direction imho.
Well, there is always some problems with learning ais...
If the objective has too many degrees of freedom (like the endless possibilities in an RTS), a sensible convergence speed for a genetic algorithm would need ages (i.e. 1000s of samples), while a too fast/radical approach would end of driving the ai into dead ends (like getting a favorite units->only build that unit->even if it sucks, others cant gain ranks because they arent/are rarely build, so even the bad unit gets more cred...)
Its very difficult to get the balance right, plus most people dont WANT to face an ai that uses only the 1 or 2 best units...
If the objective has too many degrees of freedom (like the endless possibilities in an RTS), a sensible convergence speed for a genetic algorithm would need ages (i.e. 1000s of samples), while a too fast/radical approach would end of driving the ai into dead ends (like getting a favorite units->only build that unit->even if it sucks, others cant gain ranks because they arent/are rarely build, so even the bad unit gets more cred...)
Its very difficult to get the balance right, plus most people dont WANT to face an ai that uses only the 1 or 2 best units...
Well look at what Will Wright is doing with the announced Spore and did with the Sims, there is quite a lot of adapting/learning AI in those. FPS is a different story though, because the opponents die too soon to learn.
In general, when game elements can be improved and the consumer PC's have the power for it, they will, it's just that so far not all game types were suitable for it and better behavior in them could be simulated with a simple fixed state machine.
But your statement says that "they used to learn", and now they have stopped doing that, which is reverse. Just a handful of games use it now, but it is a new trend and it'll only increase in the next few years.
In general, when game elements can be improved and the consumer PC's have the power for it, they will, it's just that so far not all game types were suitable for it and better behavior in them could be simulated with a simple fixed state machine.
But your statement says that "they used to learn", and now they have stopped doing that, which is reverse. Just a handful of games use it now, but it is a new trend and it'll only increase in the next few years.
cause there wouldnt be competition and the AI wont be the way they have planned (cause every one has their own ideas)... its better when everyone makes their own and other copy the best featuresCaydr wrote:Kick ass.
I just wonder why there's all this competition. Each AI project seems to have its own strengths and weaknesses, why don't you all just cooperate and we can say that Spring has one of the best AI codes around!
Because all the different AI projects take different routes and have different structural aspects as to how they work.
However from what i see I predict that in 2 months only NTAI and AAI will be capable of surviving, all other AI's will be crushed beneath the weight of the two.
Other than me I see nobody save submarine who has deviated from conventionality or has great plans to deviate. Look at AAI, it had basic experience programmed in before it could actualy build anything. That shows just how different submarine could be, and it's a promising sign that any AI dev who ignores AAI is gonna get hurt.
However from what i see I predict that in 2 months only NTAI and AAI will be capable of surviving, all other AI's will be crushed beneath the weight of the two.
Other than me I see nobody save submarine who has deviated from conventionality or has great plans to deviate. Look at AAI, it had basic experience programmed in before it could actualy build anything. That shows just how different submarine could be, and it's a promising sign that any AI dev who ignores AAI is gonna get hurt.