DeepMind & Starcarft AI

[riadsala]

I thought some other people here might be interested in this:

https://deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii-ai-research-environment/

[eRonin]

I'm very much looking forward to this! I think I read that the AI will be arbitrarily limited to 180 effective APM, so that it doesn't just balloon out of control in terms of its micro-management of single units

[cornchip]

It's an interesting project. @eRonin, I wonder if limiting APM even further would force it to develop better strategies? Maybe this is proof that APM matters too much in SC.

[riadsala]

Did you guys see that a different AI has recently beaten pro-DOTA2 players?

As expected, now that we have come up with a method than work well for Go (which has been a huge goal in AI since Deep Blue and chess in the mid 90s) the floodgates are opening.

[sclpls]

While the AI for the DOTA bot is very impressive, it is worth pointing out that it was just a 1v1 mid matchup with some artificial constraints involved. That being said there are some neat anecdotes about this bot:

 

-The bot eventually learned to exploit a fog-of-war rule that the AI creators were not even aware of.

-Arteezy mentioned that in one of his matchups the bot used a strategy that he had only ever seen one player ever use. So the AI has figured out some unusual strategies.

 

It sounds like their goal now is to have a working team of 5 bots by next year. Should be exciting to see the results!

[eRonin]

9 hours ago, cornchip said:

It's an interesting project. @eRonin, I wonder if limiting APM even further would force it to develop better strategies? Maybe this is proof that APM matters too much in SC.

 

APM matters as much as being physically fit matters for an athlete. Say you made a robot with AI that can play football, but then you make that robot able to sprint at 80km/h. It would be as though you allowed this SC2 AI to have way more APM than the fastest pro players.

[cornchip]

1 hour ago, eRonin said:

 

APM matters as much as being physically fit matters for an athlete. Say you made a robot with AI that can play football, but then you make that robot able to sprint at 80km/h. It would be as though you allowed this SC2 AI to have way more APM than the fastest pro players.

 

For sure. I just wondered if limiting the APM on the bot even further would force it to develop some interesting strategies. To use your analogy, it would be like forcing a football robot to run slower than humans so it had to try to win with a superior playbook, juking, etc.

[eRonin]

I did want to expand a bit and touch on what you suggested, but then I actually don't know anything about football and decided to stop it there, haha. There are StarCraft pro players out there who have lower than average APM (well, they may have since retired) and they do have different playstyles that is more dependent on maintaining centre control and positioning rather than the popular (and popularly watched!) tactic of multi-prong attacks to harass enemies in multiple locations simultaneously to strain their focus.

 

The interesting thing here is that normally, the aggressor is at an advantage due to the defender needing to react to things and make a measured response according to the scale of the attack, and thus may make miscalculations under pressure that the aggressor is not subject to. An AI is far less likely to make a miscalculation in a straight up numbers game, and they will never "lose focus" the same way a human would under pressure.

[cornchip]

Interesting! As a low APM gamer (well, everything) I find myself doing similar energy conversation strategies in games. Thanks for the explanation and I'll be watching this project with interest.