TLO vs AlphaStar

@Santah Cast by Artosis & Rotterdam



Thanks, updated :)



So I think the interesting innovation here is that AlphaStar can play *at all* with machine learning concepts. Beyond that, it's multitasking simply makes it unbeatable.

Reminds me of Overmind (?) that was programmed to just do mass Muta swarm. Just not a fair game mechanically.



No, it's not fair. They explicitly state that AlphaStar plays by the same rules, but this isn't true. The AI can see everywhere at all times as long as it's not in the fog of war. If something gets spotted for even 0.0001 seconds in the deepest corner of where it has vision, it is something that will be detected and used to calculate.

That is still an immeasurably large advantage by itself, but not only that, it doesn't have to spend APM on camera positions to move units over either, or double tap a number to focus the camera onto a unit/army. Therefore "180 APM" of AlphaStar (On top of the "per-action-effectiveness") could translate well into 400-500 APM of a pro player.

It's not just that, the number that's shown is average, and it is only limited by this average number: "Don't go over X APM in a single game" so it spends practically none when it doesn't have to (at the start) and then go 500-800 of _very precise actions_ during stalker blink to get to that average. And then call it "No this is a totally human level of APM".

Don't forget that humans also use keybinds to set up- and move the camera to buildings to queue up stuff, even during battles. AlphaStar can both attack and defend with its "full map peripheral vision" without spending this APM on the camera.

Don't get me wrong, it's wonderful to see how much Deepmind has done with neural networking, but until an AI can beat a professional player with exactly the same rules from a strategic and tactical position (instead of an APM/vision "cheat"), it isn't smarter, it's just faster.

new comment
Only registered and logged in users can post comments.