My last article for awhile: <https://towardsdatasc...

# datascience

[thomasnield]

My last article for awhile: https://towardsdatascience.com/how-it-feels-to-learn-data-science-in-2019-6ee688498029

[dalexander]

And definitely do not listen to language hipsters, like that one guy who is always talking about Kotlin. 🧌 Although I found mentioning Starcraft-trained NN kind of …interesting? Simply because of how questionable the victory was (basically proving that inhuman APM can beat human strategy, which is already well known).

[thomasnield]

@dalexander haha, I don't quite understand the StarCraft parameters either. From what I understand, it sees the entire game map which human players cannot? Does that mean fog of war was removed too?

[dalexander]

So I read a little bit about it last week, and essentially the winning version of the AI sees the entire map at once (although still with fog of war) and interacts via an API or something to that effect so it can perform multiple actions over different areas at the same time. ie. it was doing heavy micromanagement of its stalker units in 3 different screens at the same time. And while they limit APM, it’s pretty generous and it allows bursting to very high instantaneous APM.
At some point the team went back and made some changes to how the AI worked, so that it could only “See” a window like players do, and it got beat pretty badly in that game by the human catching it in a loop. The human would use a dropship to put some units on its supply line, which would cause the AI to pull its army back, so the human player would retreat, and the AI would move its army towards the front of the base, and the human would just drop on the AI’s supply line again.

[dalexander]

It’s interesting (but not that surprising in some ways) how much harder it is to take a limited window of input and build a “bigger picture” out of it, which is something the AI had difficulty with. I would also be really interested in seeing what the AI could do if it was trained to use image recognition to interact with the game rather than an API.

[dalexander]

And to give a little bit more context, I saw this video linked in the context of “This is what super-human APM can do”

https://www.youtube.com/watch?v=IKVFZ28ybQs▾

[dalexander]

But this is a pretty big digression on a single sentence in your article 🙂

[thomasnield]

@dalexander The image recognition would just add a crazy layer of unnecessary complexity imo, unless there is some academic reason for it.

[thomasnield]

But it's pretty interesting how game AI is often equated with more ambitious ideas of AI

[thomasnield]

I may write about game AI some other time. it's interesting how AI researchers always go to games, but can't seem to find real-life applications to research. I've wanted to explore why more.

[dalexander]

I think we’re more likely to get interesting more widely applicable research as we move to real time games with a continuous spectrum of moves rather than discrete moves. ie. I think research into go and chess would be much harder to apply because there are a finite number of choices. And I think the most interesting kind of information we could get out of a Starcraft AI is a new, superior strategy that humans hadn’t devised yet.

As for image processing, I think it would be interesting because it involves identify and interacting with the game the same way a human would, rather than with more accurate information. I don’t know how good we are at using image processing to categorize animated units and transforming that information into something the “behavior AI” would act upon, maybe that’s a solved problem though?

[dalexander]

It might also be necessary so that the AI could process minimap information “fairly”? I don’t know enough about how it works though, so it might already be fair.