On January 25 2019 06:30 CoupdeBoule wrote: Started watching. 30 sec into the first replay it shows that the AI isnt worker-pairing. Oh well i wasnt expecting much but im still disappointed - not gonna waste my time watching that garbage
Games 1-5 are against "basic" (GM?) agents, with one week of training. Games 6-10 are against "advanced" (top-pro?) agents, with two weeks of training but no need to control the camera - their attention is at the same time on the whole map that is visible (although their actions are not). Game 11 is against a "basic" agent with camera control.
I thought this AI play was amazing, and nothing even remotely like any AI we have ever seen before. It is playing real legit top grand master level play without any form of cheating. Its playing different strats, and its adapting very well. It has zero hard coded rules or glitches where some IF statement is missing on what to do. It never retreated and tried attacking again at stupid times. It played super well rounded solid fair games. No ai so far is even remotely close to this. Its very amazing that it learned all of this just on its own.
On January 25 2019 10:27 mishimaBeef wrote: anyone watch the game mana mentioned where alphastar did a proxy?
yes, mana threw the game and got baited hard by shield batteries.
I think Mana won the last game because the A.I refused to engage his army comp since the predicted outcome was a loss nearly always, what it didn't consider was that it was playing vs a human with flawed mechanics which it probably could outmicro even in that situation.
It has an unfair advantage with several games where it had full map view rather than camera control. It's an unfair advantage. All of the games should've been camera control, flat out.
On January 25 2019 13:29 Rea-Rea wrote: It has an unfair advantage with several games where it had full map view rather than camera control. It's an unfair advantage. All of the games should've been camera control, flat out.
Full map view as in vision of the whole map with no fog of war?
On January 25 2019 13:29 Rea-Rea wrote: It has an unfair advantage with several games where it had full map view rather than camera control. It's an unfair advantage. All of the games should've been camera control, flat out.
Full map view as in vision of the whole map with no fog of war?
No, its still limited by fog of war but capable of zooming out completely to where it can see the whole map but issue commands with perfect precision.
On January 25 2019 13:29 Rea-Rea wrote: It has an unfair advantage with several games where it had full map view rather than camera control. It's an unfair advantage. All of the games should've been camera control, flat out.
Full map view as in vision of the whole map with no fog of war?
No, the first 10 games it had a zoomed out view of the map instead of having to use a mini map where you don’t know what showed up until you move your camera to check, but the fog of war was still there. It still had to move the camera to a location to perform an action, but it was an advantage. The 11th game, it had to use a regular mini map instead of getting a super zoom out. It slightly reduces multi prong micro potential.
i watched the games the micro didn't seem undoable with good hotkey use, e.g. reassigning hurt stalkers to a quick retreat hotkey, having 3 groups of stalkers on hotkeys and double tapping around
On January 25 2019 15:01 mishimaBeef wrote: i watched the games the micro didn't seem undoable with good hotkey use, e.g. reassigning hurt stalkers to a quick retreat hotkey, having 3 groups of stalkers on hotkeys and double tapping around
Not sure we watched the same games if you think any human could do the stalker and disruptor micro the AI was pulling. It was literally microing 6+ units at a time and had its apm spiking to 1500 sometimes.
What the DeepMind team pretended not to notice. The AI uses APM as a resource, keeping it low when possible, and bursting into a superhuman micro machine when critical. It learned that overall mechanics can win games more often. There are probably more strategic iterations, but they lose to the efficient action versions (obviously).
Is there a VOD available? A friend of mine missed it.
I think the games were really entertaining. I also think #11 is the way to go, in terms of fairness. It will be nice once all races will be shown, eventually.
On January 25 2019 17:10 valas991 wrote: Is there a VOD available? A friend of mine missed it.
I think the games were really entertaining. I also think #11 is the way to go, in terms of fairness. It will be nice once all races will be shown, eventually.
On January 25 2019 15:01 mishimaBeef wrote: i watched the games the micro didn't seem undoable with good hotkey use, e.g. reassigning hurt stalkers to a quick retreat hotkey, having 3 groups of stalkers on hotkeys and double tapping around
Not sure we watched the same games if you think any human could do the stalker and disruptor micro the AI was pulling. It was literally microing 6+ units at a time and had its apm spiking to 1500 sometimes.
hmm, but if i recall, they said the reaction time was something over 300 ms and the apm graph shows this:
hmm but i guess currently it's true that the precision in its attempts to execute commands is perfect (i.e. no misclicks) would be cool to see a camera observing the screen and robotic "hand" (x-y movement) operating a mouse using control systems theory
EPT
