> A lot of it comes down to the same kind of thing. "Send your first overlord to scout their natural. If they haven't expanded to their natural, build one -- ONE -- spine crawler in your natural."

I think your general question is still a good one, but it's worth noting that from a human perspective, Starcraft matches only start off this way. Very quickly, the game state becomes complex enough that decision trees break down and intuition becomes the driving process for high-level human players.

The extent to which AI can begin to compete with this sort of intuitive human processing is most interesting to me. As it relates to this article, I think it matters a great deal if the experiment has constrained the system to the degree that it ceases to operate in that intuitive realm that high-level Starcraft matches operate in.

> Very quickly, the game state becomes complex enough that decision trees break down and intuition becomes the driving process for high-level human players. The extent to which AI can begin to compete with this sort of intuitive human processing is most interesting to me.

As far as I can tell, we're already there. Go is a game where search + explicit rules work really poorly, and intuition is a heavy factor in how humans play. AlphaGo was the first major success where neural nets were able to replicate this "intuition", and actually become far better at it than humans.

You can definitely see this "intuition" factor if you watch the AlphaStar games. Basically, if you have two armies with "middle of the road" units, AlphaStar always knows when to take a fight. You can sometimes see it take its army up to a defensive line, then turn around and go home; but if it decides to attack up that ramp, it always wins, even in a situation where a human would be less sure.

The exception is with unusual unit compositions; for example, it seems in training to have only encountered 1-2 battle cruisers at a time. So there are games where over and over again it throws an expensive army full-throttle into a group of 10 battle cruisers, and over and over gets slaughtered.

But again, I think this "intuition" factor, even in humans, is not much more than a sophisticated pattern matching, perhaps with background search.

> The exception is with unusual unit compositions

Could be wrong, but I believe this is actually a case-in-point moment for the other side. AI as it exists now is flummoxed by unfamiliar input, whereas a human expert is more likely to be able to react in a way that is at least directionally correct.

Put another way, if AlphaStar is confronted with a Terran opponent that can build Zerg units, how likely would it be to win the game? How likely would an expert human be to win that game?

No, what AlphaGo does should never be considered "intuition". It is simply using perfect information to optimize its probability of winning. We have no idea what intuition is or how it works. That's why we have learned literally nothing from AlphaGo. In fact, players who have played against it have often gotten worse. Intuition is very obviously much more than pattern matching. You need thought and probably some form of language to employ it, both of which we have very little understanding of.

The korean fight it beat human For many games afterwards.

Some insight gained like going to three three which is a taboo like old day playing three three alone.

Intuition or not ... learning we from computer go.

What the problem is they are perfectly known game. Air battle is not a perfect information game. That is why you need human decision.

> It is simply using perfect information to optimize its probability of winning.

Perfect information would be the knowledge of all possible outcomes of a given move. That's not even possible in chess, and is fantastically less possible in a game like go. That's why, until AlphaGo, there had never been a go computer program that had ever even come close to beating a professional go player.

Let me emphasize that: As of early 2015, nearly 20 years after Deep Blue beat the world champion Kasparov at chess, there had never been a computer go program that had come close to beating a professional go player. The game starts with literally 361 possible moves, and each move thereafter decreases the possible moves by exactly one. The search space is just so massive that it's impossible, by brute force computation, to do a search with any meaning.

What's needed is an intuition about two things. First, for a given a board position, which color is more likely to win? Secondly, given the hundreds of moves you could make, which ones are the best ones to explore? Go players have some rules, but primarily they develop an intuition for these two things by playing and looking at hundreds and hundreds of games.

AlphaGo's primary architecture was the same. It has two neural nets, which spit out board evaluations and move suggestions. These nets don't have explicitly coded rules, but were developed by playing millions of games. How is that any different than our wetware neural networks?

> In fact, players who have played against it have often gotten worse.

I hadn't heard this, but this is probably not unusual. Take my StarCraft example: Suppose someone had gotten to Diamond league mainly by perfecting their early rush strategies, but then hit a wall. Then they watch that video series which says the foundation of a GrandMaster strategy is having a solid economy first. If they decide to take this advice, it will entail a complete re-building of their skillset; they'll almost certainly drop in their rankings before picking back up again.

You could imagine the same thing happening in go: for hundreds of years, certain principles have been believed to be true. AlphaGo regularly violates these principles. As people are exploring this alternate strategies, they will inevitably get the "new" principles wrong while they're learning.

On the other hand, the early versions of AlphaStar clearly dominated mainly by having inhuman micro capabilities (ability to precisely control individual units). A human trying to replicate a strategy that relied on inhuman micro would inevitably fail.

Similarly, it might be that certain moves in go are good moves if you can do the kind of massively deep search that only AlphaGo can do. If that's the case, then of course humans trying to imitate AlphaGo are going to fail. A similar thing happened in chess: computers historically have been amazing at tactics and weak on strategy. A human playing a computer should focus on a good strategy, because they're never going to beat a computer at tactics.

But none of that changes the fact that what AlphaGo is doing, with two neural nets that spit out answers based not on explicit rules but based on millions of games worth of experience, is indistinguishable from human intuition.

I think you need to look up the definition of "perfect information games". It is not what yoU asume it to be, when you say "knowing all possible outcomes".

Clearly it can surpass peak human skill at StarCraft, but it’s important to understand what Deep Mind is doing differently.

One example is memory. Deep mind doesn’t update it’s strategy when playing the same player repeatedly or even over the same game. It can operate at near peak human performance indefinitely, but it avoids being exploited via deep understanding of the exact rules in play. It was also playing on the ladder under random names to avoid people developing specific counters.

> One example is memory. Deep mind doesn’t update it’s strategy when playing the same player repeatedly or even over the same game. It can operate at near peak human performance indefinitely but it avoids being exploited via deep understanding of the exact rules in play. It was also playing on the ladder under random names to avoid people developing specific counters.

This is because many games without perfect information have "rock-paper-scissor" situations. An AI designed without this in mind, lacking a source of randomness (or memory of the last encounter fed into the next game), is going to pick "rock" every time and start losing to the human who realizes and exploits this.

> Is this really that different than what humans do?

Yes, we have built in mechanism that react to situations we never encountered before. You will never be able to simply sit calmly while something is damaging you for example.

Reacting to damage is something you learned in previous generations. You've encountered that situation billions of times (you just don't remember them).

Imagination plays a significant role for us humans. We also are capable of maintaining incomplete conceptual mental models. Some of us experience epiphanies. Etc.

Pattern matching and search are brute force approximation of actual “intelligence”.

As far as I understand, AlphaZero had three main components:

1. A model of how Go worked, constructed by humans

2. A neural net for generating interesting moves, and a neural net for evaluating game positions.

3. A Markov search algorithm that used 1+2 to do pattern patching and search.

I will say that so far, #1 -- constructing mental models which can then be searched -- is something I haven't seen an AI do yet.

But looking at the actual work that I do -- from coding to deciding what to cook for dinner to deciding how to lay out my extension -- I'm pretty sure that behind the scenes it's exactly the same as 1+2+3 above; and that if I could download my mental model and my utility evaluator into a program, it could (with training) do all of the above faster and more consistently.

After AlphaZero, DeepMind went on to develop MuZero (https://en.wikipedia.org/wiki/MuZero, paper: https://arxiv.org/abs/1911.08265), which is able to learn to play games without being provided an explicit model of their rules. "When evaluated on Go, chess and shogi, without any knowledge of the game rules, MuZero matched the superhuman performance of the AlphaZero algorithm that was supplied with the game rules."

We are quite different as we can evaluate new situations based on our understanding of the principles behind something like a build order in Starcraft.

The “one” spine crawler thing is a bit overstated in the context of an actual game. It’s not like it’s 100% necessary to build it, build it in your natural, or only build one depending on what you’ve scouted, the opponent’s race, and who you’re playing against.

Like the sibling said, the game begins like a decision tree but quickly falls outside that purview.

Well you have to remember who he's talking to. You can sort of tell that lots of people, if they think they're about to be rushed, would make loads of spine crawlers as a defense; the emphasis on "ONE spine crawler" to delay an attacking army just long enough that you can build your own army in response (assuming you see it when it starts to cross the map) with the minimal impact on your economy.

And my point is that basically what he does in the series is start with a super simple set of rules, and then slowly add more subtleties and sub-clauses. That's obviously at some level how he works when he's actually playing -- he has a general plan, then a whole load of patterns that he matches on in how to respond. I'd be willing to bet that he rarely encounters a genuinely new situation for which there is absolutely zero precedent -- absolutely nothing in his experience which is similar enough for him to take an existing response and adapt it for that purpose. How is that different to what AlphaStar is doing?

The main difference is likely to be in how they learn afterwards. If he encounters a new situation in a game, afterwards he will use his model of the game to try to come up with new responses. AlphaStar will most likely have to do something much closer to brute force to find an appropriate response.