Protoss Alphastar was fantastic. It was definitely not just out-microing opponents (after later refinements Alphastar's micro limitations were very human-like), but really had an understanding of build orders, positioning, choke points and the like. Here's a great game it played against Serral (one of if not the top Zerg player in the world) where it holds fantastically against Serral's aggression and is able to use the layout of its own base to pick apart what looked like a winning position for Serral: https://www.youtube.com/watch?v=C6qmPNyKRGw
That's something which was mentioned when it came to deep mind team talking of all the "different agents" they had trained
AlphaStar's training process consisted of two phases: an imitation-of-human-players learning phase that created a base set of "templates" and then self-play among agents seeded from those templates.
When creating those templates, they basically ended up with "baked-in" build orders. But self-play among agents ended up creating a variety of reactive strategies on top of those build orders, so that agents would modify them on-the-fly as the game situation changed. At a high level this is basically how humans use build orders as well.
Both of those (and more) are filtered by EAPM measures (effective APM) and even on those metrics Alphastar still had a lower EAPM count than most pros.
I mean looking at that Serral game I don't see any superhuman micro on the part of Alphastar. Just fantastic exploitation of base layout and steadily pressing every strategic weakness Serral made in his attack.
If anything on the consistency front, Alphastar's main advantages came from superhuman macro (effectively never forgetting about making sure workers are busy), but the edge there vs human pros is pretty slight.
> steadily pressing every strategic weakness Serral made in his attack
Again a very micro limited activity. Alphastar wasn’t very superhuman in that someone could get those inputs down about that quickly, they just can’t keep the pace up over time. Endurance is one of those human limits that is often ignored when talking about esports, but it’s real and huge influence on gameplay.
This is a very narrow view of AI. In fact, the limitations you are talking about apply only to deep learning (which is frequently and mistakenly confused for AI nowadays). Other well-studied branches of artificial intelligence include planning [1] and reasoning [2].
By far the biggest problem is to formalize the real world and build a realistic simulation which the AI system can play with. After that, a decent system would blow most humans out of the water.
> Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations it was building) but simply out-microed its opponent, it just had better mechanics.
Maybe, what about go and chess? The feeling there was that humans did not understand the game very well.
[1] https://en.wikipedia.org/wiki/Automated_planning_and_schedul...
Deep learning is also where all the money and attention is going, and where the revolutionary gains have been seen in the last few years. They’re pretty good at compressing huge datasets into a queriable structure and generating features for hierarchical data, but are just one component of a generalized system that could handle something like war.
> By far the biggest problem is to formalize the real world and build a realistic simulation which the AI system can play with
Strong agree! The fundamental challenge is to create digital model of the world with enough detail & accuracy that you can run optimizations on it and trust the results. This is as true for robotic aircraft as it is for autonomous cars or an e-commerce data warehouse. People need to decide what subset of the world to observe, deploy & maintain sensors to collect those observations, build information systems to organize & serve the observations, and maintain the whole thing. Then maybe you can start running AI/ML on it.
It’s still not perfect. It still can’t play the full game with no restrictions. But it is discovering new strategies that humans never did. It’s not just micro.
The AI consistently did a 'deathball push' strategy, where small early advantages snowball and they pushed hard to exploit that early advantage. One of the things that enabled that was that resources were more evenly spread across the team, allowing the heroes that were strong earlier to do something with that advantage. Competitive and casual ranked play hasn't switch to always do deathball, but there is a marked increase in the amount of resources that early-timing 'support' heroes take, they are a lot greedier than before OpenAI. (as pointed out by a sibling post) It's unclear if that is either because of lessons from OpenAI-five, or because of the balance changes (there is more resources to be had around the map these days).
The second thing that stands out is that the 1v1 OpenAI and the 5v5 OpenAI did buy a lot more of 'consumable' items, mana and HP potions basically. These were considered too inefficient before, people tended to prefer items that gave a more over-time regeneration and long-term efficiency. What the bots did was /constantly/ ship consumables over, which was kind of surprising at the time. It used that consumable resource to either force an opponent away from xp/gold resources (lane, jungle) or allow it to stay after aggression. Once pros started buying a lot more consumables, casual players quickly followed suit.
FWIW the developer of the game has also tried to fix this over the years, devising ways to get more resources to the 4th and 5th players. For example, there is a system similar to UBI where all members of the team benefit equally from "bounty runes". But till date, the bots have achieved a level of egalitarianism yet to be matched by human players.
this kinda makes sense. I'm more of a cs player, so hopefully my observations aren't totally off base. generally, a group of five humans will have some sense of who their best player is, who their worst player is, and who's good at specific roles. for instance, you might deliberately send your worst player into a bombsite first to gain info about the opponent's defense (and save your best player to react to that info). or you might send the person who's best at "entry fragging".
an egalitarian playstyle doesn't make sense when you have human players with different capabilities. you want to play to each person's strengths. if you have an algorithm controlling five heroes (or five instances of the same algorithm, idk how it works), there's no "best player" to send more resources to or dedicate to a specific role. I don't know enough about moba to say whether this is even an advantage.
none of this stuff is particularly interesting in counterstrike. a circa 1999 aimbot with a good ping will beat any human player.
This gave him an extraordinary ability to contrast live battle situations with a huge library of learned historical situations, helping him to make decisions on battle plans.
Wargames and strategy require cognition and level of contextual awareness and understanding of human motivation and intent that falls well into the realm of sci-fi AGI. No military confrontation is generic enough to run on some ML pattern-matching or travelling salesman type algorithm.
I mean just to talk about how low-level and limited AI is, when you take any modern learning system, you teach it Atari Breakout, if you adjust the panel by a bunch of pixels or rotate the screen you need to relearn or transfer-learn the game because the algorithm doesn't even understand the ontology of the game and really just manipulates pixels.
Even Deepmind's starcaft bot was stupid in this sense. It didn't understand starcraft strategy at all (which was visible in it's nonsensical unit combinations it was building) but simply out-microed its opponent, it just had better mechanics.