One of the comments that has frequently come up when discussing the StarCraft AI Competition is that bots should be APM-limited, to prevent bots from exploiting behavior that would be impossible for a human to perform. An example of this is the mutalisk control exhibited by Berkeley's Overmind bot. However, APM isn't everything, which is demonstrated in the match below between EISBot (Protoss) and Overmind (Zerg):
The APM of each bot during the match is shown in the chart above. EISBot was able to win with an average APM of 207 against a bot with 3397. While demonstrating high APM in StarCraft is impressive, it seems that a larger APM does not always equate to better performance for bots. At what point does increasing APM no longer improve performance? And is there a significant difference between effective APM and actual APM?
Oh man I always love these AI projects, they make playing against the computer so much more of an exciting idea (Blizzard AI's = :{ ).
Well I don't think a game of mass goon versus 2-hatch mutas is super indicative, it's still nice to see the Overmind dine some time (it was against a Toss no wonder ).
Simple questions: Are you able to simply tweak the normal AI (nothing too intrusive) to only fix major issues, such as staying on 3-base max, mineral-blocked islands and mineral/neutral walls, because with those gone I wouldn't feel so bad playing against the computer to kill time (and can't reach iCCup)?
Also, why hasn't anyone made some insane AI's to pull off the most orgasmic, trans-Boxer MnM micro yet seen by man? Imagine, 1 marine > 5 lurkers (is there even a Marine v Lurker AI?). Or am I missing something? In any case, just because your APM isint 340 3397 doesn't mean you're bad (says the guy with the low-ish AMP)
Nah, it just shows that with perfect play Protoss are OP in BW :p
Just kidding
A lot actually comes down to decision making and how the APM are used. For example in the first minute or so i see the overlord changing direction 10 times in 2 seconds, which is clearly useless. It seems that the Overmind Bot just has horrible decision making and strategy, while EISBot is quite good. No amount of micro can let you win with bad decision making.
Fundamentally there is no difference between a human or a bot playing. The exact same factors that determine success in a human vs human game apply to a bot vs bot game. We all know that a strong strategy and build order is more important than the difference between good and perfect execution.
On May 10 2011 03:03 bITt.mAN wrote: Simple questions: Are you able to simply tweak the normal AI (nothing too intrusive) to only fix major issues, such as staying on 3-base max, mineral-blocked islands and mineral/neutral walls, because with those gone I wouldn't feel so bad playing against the computer to kill time (and can't reach iCCup)?
To write bots that can perform actions at the unit level, it is necessary to use BWAPI. Given this configuration it is not possible to just modify the existing AI, it would be necessary to first reimplement the AI with BWAPI and then add micro capabilities.
On a different note, limiting APM may cause bots to act more like humans, but this may or may not be a desirable quality.
On May 10 2011 03:06 Morfildur wrote: No amount of micro can let you win with bad decision making.
Is this generally agreed upon? What if you have godly zergling micro, such that your opponents can never react in time? Perhaps a 4-pooling bot can eliminate the need for strategic decision making.
As for creating bots I would argue a large number of the apm used is efficient. Obviously the APM from the overmind bot comes from microing the mutaslisks which require fewer apm than controlling dragoons.
The point where increasing apm no longer benefits depends on a how fast units take to respond. For example if you attack with a mutalisk are no matter what you input in the next 10 frames will not matter. If it takes 5 frames to change direction but 3 frames to cancel the change in direction and replace that command with another one the maximum apm achievable for a mutalisk would be 1200 apm (assuming 60 fps). The best unit to demonstrate this would be the siege tank. Do this for each unit you currently have and you have the maximum efficient apm. (This is to say if your apm was distributed as efficiently as possible). You gain nothing from additional apm you lose the option of perfect micro of all your units if you go below this point. Wasted apm (such as telling a character to move to the same location over and over again) will increase your APM but not your EAPM.
You can have 10,000 APM, perfect unit control, and still lose simply because you built the wrong units, attacked at the wrong time, didn't retreat at the correct time, etc.
On May 10 2011 03:06 Morfildur wrote: No amount of micro can let you win with bad decision making.
Is this generally agreed upon? What if you have godly zergling micro, such that your opponents can never react in time? Perhaps a 4-pooling bot can eliminate the need for strategic decision making.
Depends on how bad your decision making is I think.
And as people has mentioned before, it depends really on how is the APM used, lings are harder to micro in this case because they are melee. In this video we can see the lings sometimes kept running forwards and backwards wasting APM.
A 4 pool bot can be good but probes/drones have moving shot/stacking can counter 4 pool with perfect micro IMO, scvs also have 60 hp.
On May 10 2011 03:06 Morfildur wrote: No amount of micro can let you win with bad decision making.
Is this generally agreed upon? What if you have godly zergling micro, such that your opponents can never react in time? Perhaps a 4-pooling bot can eliminate the need for strategic decision making.
Well, if you only build Zerglings, the opponent can win the game with a single scout. You need at least adequate decision making to win a game, and that decision making is the thing that humans can do very easily but bots will always struggle with it.
I heard someone say that in SC2 (ah yes it was the "tricks with pathing" video) if you click more times your unit won't move as quickly to it's destination. My heart, died. I had always felt so pro spam-clicking on the same spot, and it was painful to hear that wasn't only useless, it was actually detrimental.
Is this true with SC:BW unit AI?
And another thing, heyoo Mr. djsherman OP, you sound capable, could you please make an AI that teaches marines to micro versus lurkers (that plus microing the heals of medics to get their health up JUST to the right level to no die in one hit). I'd really love to see a perfect, and I mean perfect marine split versus like 5 lurkers. My life would be complete
I believe when people say the AIs should be APM-limited, they mean that the ideal would be using something like AutoHotkey to manipulate the actual interface with mouse and keyboard commands instead of having hooks directly into the game's memory.
Of course, something like that would be inordinately harder to program, but simultaneously much more realistic. It would also cleanly allow the use of "proper" mutalisk micro by using control groups.
A similar project was started in SC2 using ScreenAPI to manipulate the game interface rather than through something like Galaxy Editor scripting.
I was one of those that thought APM needed to be limited, but have since been enlightened on the subject with some convincing statistics and don't feel its a concern. I wouldn't worry too much about the APM at this juncture, but continue to refine the project and the bot's intelligence.
Looking forward to seeing the project move forward with spell usage now. The micro management of units is godly and has been exciting to watch be fine tuned. Really want to see the bots start using those epic Starcraft spells now. Lockdown, darkswarm, emp, stasis, recall, consume, disruptions web, hallucination, perfect storms, blah, blah, blah.