On April 07 2011 18:02 Assault_1 wrote:
doesn't this require knowledge of the future? how does it know which ling the tank will target

awesome vid though

The tanks/banelings target their target before the actual attack so the ai detects which zergling is targeted and split accordingly

Dezire

Netherlands640 Posts

April 10 2011 12:31 GMT

#502

holy shit, mvps spread was sick

XiaN

Germany162 Posts

April 10 2011 12:35 GMT

#503

On April 08 2011 09:30 frucisky wrote:

Show nested quote +

This is a great idea. I second this

Nothing is impossible, when you use the Automaton 2000.
Brilliant job by the way!! The marine split was especially mind blowing to watch.

Another idea for you would be to micro cracking vs zealots attacking and pulling away before the zealot can swipe for damage. Don't know how possible this is..

This idea sounds fun ( Stalker vs Stimmed Rauders ).

Also : Kill an Ultralist with Drones/Probes/SCVs by avoiding the Splash damage with the help of Automaton :D

InsaniaK

Sweden120 Posts

April 10 2011 12:54 GMT

#504

On April 08 2011 02:29 Elean wrote:
In the zerglings vs tanks, it's not micro, your bot is cheating by using the info of which ling is targetted. Even with perfect micro, it is not possible to achieve with normal information.

Marine vs bannelings is different, it really shows the limit of the unit. But to be fair, the bannelings sould also be controlled by a bot.

Is it really THAT hard to read?
It has been stated multiple times that it doesn't cheat to see which ling is being targeted. It just predicts it and according to the OP it is rather easy to do so.

Katagiri

Lithuania3 Posts

April 10 2011 13:36 GMT

#505

Would be very interesting to see what could this BOT do with a limited amount of APM. I mean if you would give this bot an APM limit, one could get a lot of insight into important parts of micro. For example: could the program achieve same survival ratio with 250 APM?

I assume that right now the program just does what it has to do in order to achieve the best possible results, but if it had to cut unnecessary actions or less important actions to achieve same results, it could potentially find the key actions required for micro. It could even find something that SC2 Pro's could incorporate in their game.

All in all, I am super curious what could this BOT achieve with low APM.

Bobster

Germany3075 Posts

April 10 2011 15:27 GMT

#506

15000 APM :lol :lol

Fun stuff, really cool to watch.

AbeToss

United States60 Posts

April 10 2011 15:40 GMT

#507

On April 10 2011 22:36 Katagiri wrote:
Would be very interesting to see what could this BOT do with a limited amount of APM. I mean if you would give this bot an APM limit, one could get a lot of insight into important parts of micro. For example: could the program achieve same survival ratio with 250 APM?

I assume that right now the program just does what it has to do in order to achieve the best possible results, but if it had to cut unnecessary actions or less important actions to achieve same results, it could potentially find the key actions required for micro. It could even find something that SC2 Pro's could incorporate in their game.

All in all, I am super curious what could this BOT achieve with low APM.

I'm also curious to see how a bot would react with limited APM or in simple scenarios encountered frequently in game such as the probe/zealot/stalker encounter in a PvP. I feel like we could learn a lot by watching a micro bot.

It's be cool to see the bot use perfect FF's. MC force field donuts anyone??

Kajor

Singapore5 Posts

April 10 2011 15:50 GMT

#508

put blinkstalker with blink micro against each other. AI WARS!!
It'll be fun to watch.

deadjon

United States83 Posts

April 10 2011 15:54 GMT

#509

On April 11 2011 00:40 AbeToss wrote:

Show nested quote +

I agree here 100%. Playing out scenarios, with a variety of APMs could produce some very interesting results.

VictorX

United States773 Posts

April 10 2011 16:00 GMT

#510

The automaton's apm was over 9000? How can we lowly humans even compete..

Have you ever considered crafting a 4v4 comp stomp where the cpu is given this sort of insane AI? It would be amazing

gm.tOSS

Germany898 Posts

April 10 2011 16:01 GMT

#511

On April 11 2011 00:54 deadjon wrote:

Show nested quote +

I agree here 100%. Playing out scenarios, with a variety of APMs could produce some very interesting results.

Might be hard to teach the ai how to priorize the commands.

Without priorization the ai would look plenty stupid with 250 apm probably.

Pulimuli

Sweden2766 Posts

April 10 2011 16:01 GMT

#512

the marine splitting made me laugh in awe :D

MrCeeJ

United Kingdom57 Posts

April 10 2011 16:11 GMT

#513

Although not a starcraft pro I can give you some insight into the problem of limiting apm from an AI coders perspective ..

The bot at the moment can simply read in the targets of the tanks (or work them out itself) and then generate the list of doomed lings, and simply send a move command to all of the nearby lings to move away.

It moves not only the lings under the splash effect, but also the lings around those lings to make space. Essentially issuing move / attack commands to all of the lings multiple times a second.

If you limit the APM, you are constraining the output of the bot to a much smaller number of commands, so instead of just executing the logically - deduced 'perfect micro' you have to create instead an 'as good a micro as I can with only x100 APM'. This kind of optimisation problem is considerably harder than a solving problem. For example comparing the aiming of a gun in a fps (point at the head correcting for movement) to playing a game of chess (choosing a move given a large amount of possibilities and consequences) gives you an idea in the jump in complexity.

In order to perform the calculations necessary to reduce the perfect tank micro to a list of just some 200 move commands per second the simulations would need to be done incredibly quickly and is probably way out of the realms of possibility of galaxy code running on a high end machine / small cluster.

Even just deducing the minimum apm - if you have a tank hitting ~15 zerglings you need 14 move commands just to save them from the splash, assuming perfect planning that is another 14 moves to make room for them. You then need to correct all of these moves so they go back to trying to hit tanks, so you are already up to 56 move+attack commands to dodge one tank splash. assuming a 'sensible' apm limit of 500 you can only avoid a couple of tank shots a second assuming that EVERY ling not directly affected by the splash is given no commands at all.

It might be possible for a apm-limited bot to perfectly micro against a small number of tanks (3-4) but any higher and you have to accept a non-perfect (i.e. optimised) set of instructions, and you run into the computation complexity problems mentioned above.

I hope this wasn't too geeky but a few people seemed interested in the AI side of the consequences of limiting, so I thought I would share

Kazang

578 Posts

April 10 2011 17:31 GMT

#514

On April 11 2011 01:11 MrCeeJ wrote:
Even just deducing the minimum apm - if you have a tank hitting ~15 zerglings you need 14 move commands just to save them from the splash, assuming perfect planning that is another 14 moves to make room for them. You then need to correct all of these moves so they go back to trying to hit tanks, so you are already up to 56 move+attack commands to dodge one tank splash. assuming a 'sensible' apm limit of 500 you can only avoid a couple of tank shots a second assuming that EVERY ling not directly affected by the splash is given no commands at all.

To be fair you could have the AI move a group of lings as one "action", which would mimic (to some extent) the human ability to box select groups of units.

So even with an apm limit it would get pretty good spread on the zerglings with minimal actions, more apm would allow smaller groups of units to be given the same command scaling to each zergling individually getting it's own separate move command as the apm limit rises.

It would give some variance and artificial noise to the spreads, still un-humanly fast but with reasonable flaws.

kaztal

Sweden68 Posts

April 10 2011 17:57 GMT

#515

banshee vs marines, no stim
hydra (with range upgrade, on creep) vs something. reapers?
automaton cracklings vs automaton blue flame hellions

KillerPlague

United States1386 Posts

April 10 2011 18:02 GMT

#516

i'm thinking pull a spanishiwa style and have mass queens transfusing vs something buff. perhaps voids? or is that alrdy one sided in queens favor? maybe carriers then :D

Jesushooves

Canada553 Posts

April 10 2011 19:16 GMT

#517

Excellent first post, a very well edited video showing that maybe there is no skill cap to sc2

at least micro wise.

On related to op note, burrow micro roaches would be interesting, while we commonly see blink micro, burrow micro tends to be neglected, I would like to see it!

Tachi

Norway28 Posts

April 10 2011 19:38 GMT

#518

Use the blink away from missile trick and make 100 stalkers vs 100 marauders with no stalkers taken damage ^^

Naughty

United States114 Posts

April 10 2011 19:47 GMT

#519

I Would like to see zerglings vs a full protoss death ball.

phANT1m

South Africa535 Posts

April 10 2011 19:53 GMT

#520

OMFG, that Marine micro was freaking crazy. The zergling thing not possible at all as a play will never know what is going to be targeted but the marine might be what future koreans are :D.