Serral deliberately played absurdly straight up and having spoken to him the bot just micro'd at 10000 apm and killed him with a few units very early. It did not do anything interesting or allow any games to progress in any way. Serral only lost because he refused to abuse it
-
-
En réponse à @nathanias @DeepMindAI
Ah ok. That said, the weaknesses it has now are a product of what it's been learning from. There's little reason to expect it to not learn to cover those weaknesses if it were exposed to more exploiters. Regardless, it would be cool to see a Man vs Machine sort of event.
1 réponse 0 Retweet 0 j'aime -
I haven't seen a single agent that understood scouting properly - it's narrow intelligence seems to be unable to solve the strategic elements of Starcraft. Perhaps it needs more processing power to hit more billions of practice games - or perhaps SC is too complicated for it
1 réponse 0 Retweet 2 j'aime -
That's a really interesting observation. Could you elaborate on what you've seen that makes you think it doesn't understand scouting properly? (I haven't watched enough DeepMind matches to really speak to this and even if I did I'm not sure I'm good enough to do so)
1 réponse 0 Retweet 1 j'aime -
Each agent repeats a generic strategy over and over and does very little scouting/map vision. It doesn't scout and react very much at all. Most of its reactions are purely tactical (how it moves and controls its units)
1 réponse 0 Retweet 1 j'aime -
Which is insanely difficult to learn which is cool - but it doesn't seem to have figured out the value of scouting and adapting. It plays solid by using generically good unit comps rather than adapting/reacting
1 réponse 0 Retweet 1 j'aime -
Ahhh okay. That makes sense. I wonder how much of that is due to how it was told to "win". As I guess if you think about it, there are two different ways to optimize for winning: having a generic strategy you apply against everything or having strategies for specific situations
1 réponse 0 Retweet 1 j'aime -
Part of me wonders if the ML decided somewhere along the line that the former was better or at least more or less the same as the later. I'm also a bit curious how much of this is a limitation of what it can "recall" as they talked about how it can end up chasing its own tail
1 réponse 0 Retweet 1 j'aime -
As it's largely optimized to be good against what it's been playing against which means it could can possibly end up becoming to an old strategy it was once not vulnerable too. Either way, I do hope they do more with it. It would be sad if they left it at this and moved on :(
1 réponse 0 Retweet 1 j'aime -
Also thanks for the explanation!
1 réponse 0 Retweet 1 j'aime
Yeah i think it's such a daunting step though especially with a game with imperfect information and complex game "moves" the complexity of SC over Go is exponential - i hope they do more!
22:34 - 3 nov. 2019
0 réponse
0 Retweet
1 j'aime
Le chargement semble prendre du temps.
Twitter est peut-être en surcapacité ou rencontre momentanément un incident. Réessayez ou rendez-vous sur la page Twitter Status pour plus d'informations.