That doesn’t really sound like it understands the game... is this not training by playing against itself? How is it not learning the flawed reasoning behind this strategy?
The problem is that there's a sort of rock-paper-scissors game of (conservative)-(greedy)-(all in). That is, a player who build defenses early on will crush a worker rush but they'll be far behind a player who ignores defenses and focuses on growing their economy. The worker rush is designed to punish this greedy build order.
There's nothing that can be done, really. If you decide to send a worker to scout super early on then your economy suffers so much that you'll be behind no matter what you do.
Edit: to those of you drawing a distinction between an immediate worker rush and something like an early pool or a cannon rush, that's splitting hairs. These are all classified as all-in strategies that almost automatically lose if they encounter a conservative build order.
The term means you take your initial workers and charge the enemy. If you know where the enemy is you'll arrive about the time they are up one or two workers on you, then you just have to fight when you're down one or two workers and you can't back off (they'll return to mine).
A human can win like this if they are so much better at micro that they can overcome the material disadvantage (unlikely) or if their opponent panicks and does something dumb (e.g. pulling only half their workers to fight).
Because worker rushes arrive so early it doesn't matter what plan you have for the match, you'll have at least as many workers (if you are also worker rushing) or you'll have more. Maybe if you're going six pool you'll have fewer but you'll also be able to get lings and win with them.
The computer can lose to worker rushing only because it can get caught in loops of poor decisions (e.g. you lure away a couple workers and kill them).
Worker rushing is always a bad plan.
This is common technique, mostly used to get scouting drone back through a walloff (own unit stands on "hold" in a small choke) or when a mineral line is under siege (e.g. Siege Tank, Liberator), the player can select all his workers to mine a single mineral patch (has to be spammed, otherwise workers redistribute) outside of the siege range until they siege is cleared up.
As the grandparent stated, a worker rush only works against the built in starcraft computer player, because the computer AI is too dumb to recognize "I am being worker rushed" and may not commit all of its workers to defending.
Things like early pools and cannon rushes are (depending on the details of the meta at the time) a rock papers scissors game. Cheese (early pool/cannon rush) beats aggressive play beats conservative play beats cheese (roughly speaking).
Worker rushes are just bad. Apart from playing against 6 pools on really old versions of SC2 on maps with ridiculously short rush distances they just lose. Even then I don't think I'd call them good.
That might solve its task, but it’s a far cry from learning to play Starcraft well. I was surprised they seemed to have trained it against the standard Ai as that’s not useful.
If it's playing against the Starcraft 2 blizzard AI, it will learn a lot of "bad habits" because outright bad strategies work against the AI. Thus the worker rushing.
Your "reasoning" is actually a false perception. It doesn't make these mistakes and tries strategies that seem improbable to human players.
A 50% win rate against the Insane AI is very good, as at the higher AI levels the computer actually cheats (no fog of war, more resources, etc).
The AI cheats, but is also incredibly predictible, and bad at decision-making.
If you do this long enough you can build a virtually undefeatable siege tank army that the AI would simply smash its face into, and then corral them to 2 or 3 bases. They would always run out of minerals while I expanded everywhere and slowly inched tanks forward. Doing this same strategy against humans got me to about gold tier in the early days of sc2.
It's not my fault if you don't scout. Planetary Fortress rushes, proxy barracks, cannon rush... it's all fair game.
> Scholar's Mate has sometimes also been given other names in English, such as Schoolboy's Mate (which in modern English perhaps better connotes the sense of 'novice' intended by the word Scholar's) and Blitzkrieg (German for "lightning war"), meaning a quick and short engagement (Kidder 1960).
> After feeding the agent replays from real players, it started to execute standard macro-focused strategies, as well as defend against aggressive tactics such as cannon rushes.
The reason is that while the attacking player sends over their workers to attack, the defending player is still constantly making workers. By the time that attackers get there the defending player should have more workers. He simply has to command all of them to fight the opponent's workers.
If the opponent doesn't react, you win because you kill all their workers.
If the opponent overreacts, you win because you are mining more than them (IIRC this is a problem with the insane ai).
If the opponent reacts poorly in other ways, you might be able to win by out microing them. E.g. SCVs (terran workers) consistently beat drones/probe (the other races workers) 1-1 without micro. Repairing can be an advantage. The offensive AI being better at stacking damage on a few units. Etc.
The defensive AI should learn to fix these problems and stop losing to worker rushes, but it could take awhile. In other words the defensive AI starts out as a lower ranked player.
PS. I'm not sure how serious a tournament this was, but here's an example of a real worker rush (not an early pool with workers attached): https://www.youtube.com/watch?v=_AtqgaJzJP4
Of course you lose all of that time it takes for your workers to drive across the map, so it's a pretty dicey strategy. A 50% win rate against the AI sounds about right.
Or is "understanding" an attribute which is only afforded to humans who do these things, much like a machine can process the same inputs in the same ways as a human but will probably never be considered to have 'qualia'?
Using this strategy is objectively bad in that it can be repelled by an opponent almost regardless of their initial strategy.
If deepmind’s goal is to beat the standard Ai, then sure I guess it understands that. But that’s a very different problem from understanding Starcraft, and a much less difficult one at that.
And once the defense gets good enough, the attack portion of the AI should probably do the worker rushes much less often.
Of course, given that the AI has different constraints in micro-management and attention, and that the game is balanced given human skills, some things that would be totally bonkers to do as a human might work for the AI.