Hacker News new | past | comments | ask | show | jobs | submit login

This reminds me of Starcraft AI experiments. They can't actually make the computer smart, so they just jam 2000 button presses per second down the tube, giving every single unit its own simultaneous AI, and it out micromanages anyone.

With Marines usually.




I heard that the DeepMind Starcraft project intends to limit their AI's APM (actions per minute) down to something human-like.


I read that too but I hope they know the difference between APM and EPM. Pros spam APM that they could never do real actions on but their EPM is considerably lower (if the bots make actions based on pros APM they will have an insurmountable advantage).


Pros spam APM to keep warm, during battles or production macros they will frequently have a high EPM as well.


There's also accuracy.

Unlike a human, a bot will always "click" exactly where it intends to.


That's not what Starcraft AI field is about. They actually started with a combo of people doing planner-oriented systems and micro-oriented systems. Hybrids followed that. There's many methods at play. Here's a survey:

https://www.cs.mun.ca/~dchurchill/pdf/starcraft_survey.pdf

The competitions that involved humans showed humans destroyed them by spotting their patterns and beating those patterns. Also with bluffing or distractions such as having one unit do weird things around their base as the human player built up an army. The bots that beat humans will have to learn to spot bluffs and other weird patterns humans will do to screw with them. On top of all the stuff prior AI did with human-level talent. My money is on humans for DeepMind vs Starcraft although I'm happy to be proven wrong.


I don't think DeepMind will get into a high-profile competition against human pros unless they're fairly certain of winning. So if we see the equivalent of the AlphaGo versus Lee Sedol match being announced for Starcraft, then my money would be on DeepMind.


In starcraft there's a much bigger advantage since humans are inherently "single threaded", and so you can get much bigger discrepancies in APM (or EPM). Smash is more like 1 unit vs 1 unit micro. The precision and timing are still advantages for the AI, but not so much raw parallel compute.


Broodwar bots perform poorly against competent humans though. Micro advantage or not, the strategic decision making isn't there yet.


Or individual muta micro. That was the winning "strategy" in the first BWAI cup many years ago.


I remember this. And even the first bot vs bot competition was won by a clever bot scripter simply denying the gas early on from his opponent. Took all his competition down to a standstill.


Honestly, it still isn't even that good. Best startcraft AI in the world that cheats, still can't beat the low tier pros.


That's an interesting definition of smart that doesn't including being able to manage hundreds of units simultaneously.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: