A.I. too overpowered

Yes, I recall it was big news because it came out with one build that humans hadn't run across because it was pretty counter-intuitive to the thinking of the time in the way its timings worked. But this was possible because the Starcraft 2 opening game is predictable down to the second. With random maps in the PA style you couldn't recreate that.

Isn't that a bit reactive? Shouldn't the AI be expanding simply because it can? Further expansion means more factories, more factories means more units, more units means it's more likely to win. Not to mention that taking a spot is denying that spot to your opponent.

Yeah, there was some ridiculously counter intuitive rush build discovered by the program. (the five roach rush, if I recall, which swept the ladder for a week)

In Starcraft 2 the only difference between maps for the purpose of this program is the time your workers need to move to your second base, which is not too significant, this makes it very predictable. But maybe the map features in PA can be generalized or maybe they are not too significant. I wouldn't really know.

Maybe a random question, but do you think you could explain some ways you(or we, as a player) might be able to 'customize' an AI to play against later? Or will it just be settings like we've classically had in games? (Easy/Medium/Hard/Extreme/etc)?

The thing is, more fabbers = victory is not always true. At the moment, it seems to build 3 fabbers and then military. I've had starts where we're close together, and by the time I have 5 fabbers out building my metal extractors, it's destroying them with ants. It could win those games if it was more sophisticated, and without needing more fabbers. The reason it doesn't win isn't because I out economy it. It's because it doesn't take advantage of having more tempo than me, and I'm able to regain lost tempo, often just by building turrets and pelters and regaining the lost metal areas. My economy may be at 50% or worse at this stage.

It's not to do with the AI having less fabbers, it's because the ants don't maintain the pressure. Depending on how that initial attack went, the AI may have reduced me to one fabber and my commander, and in its base it may have begun going tech 2.

My point is that it needs behaviours (military, economy, production). If it ALWAYS builds 5 fabbers at the beginning, then you just build 5 bots at the beginning, start raiding it, and slow down it's initial economic development. Then you pump out fabbers, become huge and win. The thing is, you don't win because you become huge. You win in the first 10 minutes because you killed its economy, gained more tempo than it had and took full advantage of it.

I understand where the AI is up to right now, and I agree that adding more fabbers will make the AI more challenging. My point is that once the AI is more sophisticated, we shouldn't be able to predict whether or not is playing offensively, defensively, aggressively, economically, expansionist, factory-complex, rushing T2, etc. Or how many fabbers it has on the field in the first 3 minutes, build order of first factory, etc.

People's initial econboom build is very susceptible to rushing. It's just no one has perfected the art of a strong rush yet, it's made harder by the random map layout, and one well placed turret can prevent both air rushes and land rushes from doing any damage. But it's a perfectly viable strategy, and the final build of the game shouldn't be turned into "who booms faster".

That was my thought too. It doesn't expand to areas of strategic importance. But it doesn't have a strategic master plan. SO that isn't surprising.

It's kind of funny when it decides it has enough metal and just pumps out units until the planets increased mass causes it to turn into a star.

Sorian - can we give the AI a name? And a gender?

It is reassuring to know that professional software engineers make typos in code also.

At some point post release when the pressure has died off, are you able to describe the AI in detail, the approaches you used and so on? And will we be able to mod the AI a little (not for me to improve it - I'm not that great at coding. Just to read the code and see what's going on, make changes and see what happens. More experimental)

I know Don't worry. I sort of semi understand where the AI is up to. I'll get off your back.

I don't think the AI being able to cheat is a bad thing, I just don't think that should be the way to control the AI difficulty.

Right now there is a button (Greyed out) to control any player's eco handicap. When that gets implemented I'm sure some players will use it to make the AI tougher. But I in general hate it when developers try to make the AI harder by making numbers bigger. Weather it's giving the enemies more HP in FPS or more Eco in RTS, I find that gets used as a crutch when the Devs can't make a good AI, and that the end result is a game that feels cheap instead of challenging. One example that's not an RTS is Deus Ex Human Revolution. The guards are still stupid on hard mode, but now they can kill you in one hit. I won't feel particularly clever for outsmarting them, but if they do spot me I will probably die right away because there guns are magically more powerful than mine.

I'd prefer a game where the difficulty level controls how aggressive the AI is, not how much firepower they have. And I'm really excited to see Sorian's work develop.

Also I think it's kinda funny this thread started with a player complaining about the AI being too good, and we immediately change the topic on how to make it even harder.

Don't worry antonyboysx. A strong AI can easily be made dumb.

I once heard you guys mention something on the lines of actions per minute etc. So it would be equivalent to how many times a slower player clicks per minutes etc. Is this still a common ai difficulty lowering technique? I have heard that a Final AI is its most difficult form before moving on to cheating, and setting the level is based on how you detriment its abilities downwards. Is this more on the lines of how you would adjust an ai?

In theory an AI can see the entire map at once and could micro control every unit independently.
It's not limited by a screen and a mouse like the player is. And this advantage will only get worse as the game progresses as an optimized AI would be able to control units on several planets at once.

you're building this AI from most agressive down aren't you?
meaning you first make the best, then you can decline it's weaker versions.

you should probably have a talk with gunshin.

adopting the best existing strategy sometimes is beyond us, sometimes it's like learning in math about a new, second logic that you can use and it kinda breaks your brain at first.

you just reject it.
but once you can adopt the new strategy then it becomes obvious how much it was the right way to go from the start.

the point is you don't want to multiply............. you don't want to be exponential....... no you wanna be factorial. right from the begining.

Keep in mind that the goal of the designers is not to create an AI that is powerful in its own right, but rather create something that is fun to play against. Creating a strong AI is just a means to an end of not having an AI that can be figured out or predicted too easily. If sorian wanted to I'm sure he could find some micro behaviors to exploit where it's beneficial to have the AI's capability to do 2000 actions per minute, ensuring that you could never win a proper game versus the AI.

Coming from the Spring engine development mindset I strongly disagree. In most Spring games you can enable any custom script you want and that script could even be an AI playing the game for you.
The hardest AI should be as strong as possible and it should use whatever methods it can to gain an edge. If the AI can use massive APM to abuse some micro then so can a human. Such an AI would reveal such exploits for the developers and one of the goals in PA is to decrease the need of micromanagement in favour of large scale management and strategy.
Instead of banning such a script, Spring developers are more likely to patch the game to remove this micro sink or actually include the script in their game and include it in the UI so that all users can make use of it.
In my opinion microable exploits should either be removed or automated.

Can you give an example of a micro exploit? Is this like having Dox walk in circles around the enemy so they can't be hit?

Yes, that would be one example. Although a continuous stream of bullets would hit a unit moving in circles at some points and a circling means that the unit can't advance or retreat so it is still a situational tactic.
Other exploits can be placing nanoframes, starting to make buildings, to block projectiles continuously.
Massive kiting could also sort of be considered a micro a exploit so I've always said it should be automated because I think it is an interesting tactic with a lot of counterplay.

This is just silly. A human doing 300 actions per minute, with a high lack of precision, is not at all comparable to a computer doing five thousand such actions with perfect precision. That has little to do with revealing exploits, the design could be perfectly alright for human players but break under the pressure a computer AI can place upon it.

Starcraft II examples:



I'm sure there are some for PA.

mein gott (Is that spelled right? Anyone know? xD )

Just imagine if the pros could play like that, holy crap.

No. If we could do that there wouldn't be such a thing as cheating AIs, all difficulty levels would simply be handicapped versions of the master AI. They're not because we can't make unbeatable AIs in RTS. Massive APM doesn't equal victory. If it's 2000 APMing into your turrets then where's the threat?