Hacker News Reader

Comment by kebman | Hacker News Reader

You're probably surprised that you don't see it, because it's top secret.

Here's an example of something you don't hear about: AI Assisted Battlefield Tactics. While I've never heard of it (certainly not in the press), I simply assume that they're doing research on it. I base this assumption on the AlphaStar AI from DeepMind that would completely pawn Grand Masters in the game of StarCraft 2.

While AlphaStar is obviously designed for a game with very set rules and weights, I don't think it's unfeasible to develop a similar military AI that could also work in the shifting tactics of the modern battlefields.

The result would be comical, as larger armies without such an AI would always lose to someone with that AI, even if their army is much smaller or seemingly technologically disadvantaged. And it would seem that there's at least one such country in such a position today, which is also heavily supported by the NATO.

View on HN

AlphaStar hardly pawned GMs in any meaningful way. It mainly won on mechanics, even when Deepmind tried to nerf its mechanics. Probably the most important thing is that the “200 APM limit” is way higher than any human can do, because it made use of each one of its actions. Humans can reach 400+ APM by spamming, but their EPM will stay much lower than 200. The result is that it had god-tier micro that changed the game completely: zergling runby’s are normally good for Zerg and bad for Protoss. With alphastar, it was the opposite, because it could micro to avoid losing any workers. Its disruptor micro was again unmatched.

Now, my last argument for why alphastar wasn’t some sort of brilliant tactician was that eventually people figured out that it’s susceptible to all the same problems a normal bot is: you can dart a warp prism in and out of its vision and it will send its army back and forth, you can confuse it by going mass raven/PF etc. and it’s completely unable to adapt. So in conclusion, no: AlphaStar doesn’t say anything one way or another about military AIs.

The word is pwned, not pawned, by the way. I suspect your autocorrect, as well as op’s has had it’s way with your prose.

I often wonder how much autocorrect will shape the evolution of language.

s/it's/its/

> I often wonder how much autocorrect will shape the evolution of language

Apple recently announced that it was going to stop autocorrecting a certain expletive intensifier to "ducking", and I think that one actually had the potential to catch on....

I wonder if it'll ultimately do more to preserve ridiculous quirks of English spelling (and US/UK distinctions) most people don't remember or care about than change language

That’s hilarious, I did actually write “pawned” intentionally, but while writing it I thought, “Haven’t heard this one before, but I guess it makes sense, it treated them like pawns in chess”

There are plenty of niches but most people aren't aware of them: surveying and navigation. There's nothing particular secret about an RLG or an MRG at this time. There are enough examples of civilian industrial equipment using them. There are also MEMS-based accelerometers, such as those in your phone, and ones with several orders more precision and accuracy.

Applications include:

- Underground

- Under dense rainforest canopy where GNSS signals don't reach

- In narrow and deep valleys

- Salt water deeper than 200 mm

- In space, far away from GNSS systems (GNSSes work in space)

Source: I may have wandered through Trimble Nav Ltd. once or twice.

The battlefield is nothing like StarCraft though. StarCraft gives you real time control and visibility of every soldier and every enemy within the view range of any of your units. That has approximately no overlap with how command works in real militaries.

The battlefield is nothing like StarCraft? Think of SC as a top down model for how a real battlefield map can be. Just give each actor more realistic traits. So no strange Protoss or Zerg creatures! Perhaps base it on a more probabilistic model that treats each contact more as an uncertainty, like it would be under the fog of war.

Suddenly you've got a viable model to use on real tactical maps, that is strikingly similar to SC2, though adapted to real life conditions. That is, if you're able to track - or in the least theorize the position of - each individual "piece" on that board (so soldiers and armour, etc).

A real life soldier does perhaps have some different qualities and abilities than a space marine, but it's still a great starting point for building a more realistic model.

So say, your model looks more like Arma 3... Only top down. Now, I've been in the army, so I can tell you first hand that many aspects of Arma 3 is scarily realistic. Other aspects, not so much. But if you add in some modding, and all that is solved.

It's the same principle. You've already have a base AI to "play the game" and now you adapt each information point to reflect real life instead of game mechanics.

Sure, this sort of AI is brand new, so there would of course be things to iron out. Perhaps you need to use other statistical models to better capture the fog of war. But at the end of it, I think it can be done. Hell, perhaps it can even be used to fight crime. What do I know, perhaps it already is, but nobody wants to talk about it, kind of like how it was such a great moment when it dropped that New Scotland Yards used graph databases to solve cases before it was cool.

Arma 3 might be somewhat realistic, but that has nothing to do with what I’m saying about Warcraft 3. I’m saying the level of micromanagement and perfect information in W3 is completely unlike how actual battlefields can be managed.

AI that is good at W3 manages every individual unit’s actions and it does that based on a map that magically reveals enemies when you get close. Nothing like that would ever work IRL.

Acushully...

There is a very specific problem - the simulation speed problem - that is in the way.

For games RL agents can be trained by running billions of adversarial simulations. Set up a battle between two agents and reward the one that wins, try variations and repeat, and repeat, and repeat. The trick is to manage the search process so as to not waste time on strategy branches that won't get anywhere, but the idea is very simple.

Unfortunately for things like real world conflict the simulations aren't that representative of what really happens, they are also really really really really complex and slow, and so finding wining strategies is a bit hard.

Reply on news.ycombinator.com