OpenRA

I was reading recently about a Poker AI bot (in fact there are quite a few apparently). But this particular one taught itself to play poker by simply playing millions of hands at high speed and building a decision making process around it's successes and failures.

I was wondering how this might possibly work for a strategy game just like this one. Obviously the complexity is far higher but still I think theoretically possible.

Can any of the more technical people put an opinion in on this? Just how much more complex would it need to be? What challenges would it face compared to a much simpler game like poker?

Of course, I'm not actually suggesting that the devs should try and do something like this to create an incredible A.I.

I'm merely looking for a bit of almost OT discussion about the subject.

Technical aspect aside.
If you'd implement that sort of AI into a game like RA and have it learn from all its players I imagine it would get too tough to beat for a lot of players really quick. Unless ofcourse you scale difficulty but that sort of defeats the porpose.

An interesting idea however it would need to realize that it has to scout and utilize what it has scouted rather then fog of war cheating.

Its possible to have the AI register user names and have it logged as "This user likes to do this build and etc". This however, can get problematique, if the player knows multiple builds. Then of course you could find yourself with a fragged drive.

To me it seems that this could work for small tactical fights. If the computer conducts millions of small tactical battles in various setups, then it might learn that its good to take some rocketsoldiers along with artillery, to attack rocket soldiers met regular infantry, to spread artillery, etc etc etc. If it learns all that then it might get very strong after BILLIONS of trial and error runs.

For larger strategical decicions I think it is most important for the AI to "understand" and react on two things:
- what kind of map is this (shape, water, etc) and how does that influence my buildingqueu/preference? And what points on the map should I dominate and expand to?
- where is my opponent?

you guys are over thinking it all you need to do is once the ai is made have it battle it's self with one human player so like 7 ai and one human that way it learns. you need to have siad human player have more then one tactic or it will fail.

Im not an expert regarding Artificial Inteligience and selflearning computers, but as far as I know it takes some complicated and sophisticated programming to create such self learning bots. You cant just say "learn about previous game".. you have to create stuff that measures succes and failure and you got to program stuff that understands what lead to succes and failure. Again, Im not an expert, but that seems very difficult to me.

Especially when you apply it to a larger scale: entire games. What led to victory? Good microing? Good basestrategy? Bad play by the opponent?

On a smaller scale I do see oppurtunities:

For example:
5 rifle infantry vs 1 artillery.

1st try: together towards the artillery, all get killed in 2/3 shots (failure)
2nd try: with space in between towards artillery, all get killed in 5 shots (lesser failure, but failure)
3 try: split up and move around arty (out of range), then attack. Many shots hit arty, but not enough to kill it. All get killed in 10 shots. (failure, but close to succes)

AI should learn: a single arty can be attacked by infantry, best strategy is to surround out of range and then attack. 5 is to few, lets try another X runs to see how many rifle infantry is needed.

Now do it again, with 5 rocket soldiers, grenadiers, etc.
Then again against 2 arty's.
Then against moving arty
Then against Arty + hind

As you can see it will take 10s of thousands or even million of testruns to get good microstrategy. And you only get succesfull testruns if you measure succes and failure correctly.

Succes = surviving
Minor succes = maximum damage dealing
Failure = enemy survives

And even that is not always correct. Because 50 rifle infantry will probably kill an arty, but can we call it a succes if only 2 or 3 are left? So should succes be measured in money? That's an option, but in some cases the unit must be destroyed at any cost: a tanya in your base for example. She has implied costs, because if she doesnt get stopped she'll destroy half of your base (=huge costs)

Well.. those are some thoughts on this matter. Interesting but difficult stuff. And now we are only talking about MICRO. MACRO AI learning is even more complex - I presume.

I didnt publicize it but im working on something to the effect of this; im not too far into it though and i dont have much time to do it at the moment; will update on this thread in future if i get back to it any time soon

AoAGeneral wrote: ↑ An interesting idea however it would need to realize that it has to scout and utilize what it has scouted rather then fog of war cheating.

Its possible to have the AI register user names and have it logged as "This user likes to do this build and etc". This however, can get problematique, if the player knows multiple builds. Then of course you could find yourself with a fragged drive.

It wouldnt be a problem if the other player was using different build orders as it is learning what to do in specific small-scale situations; it deals with things as it finds out about them - ranking them in importance and then executing a solution that has both availability and a high chance of success; taking into account the user's previous reaction times to previous attacks. I believe the saying is "Divide and Conquer"