I was one of the 2-3 people who started BWAPI (the interface into
Starcraft1) back in ~2008. We reverse engineered the whole game and
found offst by offset ... function entry by function entry and
hooked/detoured them all. I got a very good understanding of how
"scripts" interact with Starcraft.
Back then I wrote very primitive if-then-else driven scripts which used
our API in order to allow me to win alone vs. multiple top rated players
simply by abusing the fact that I could control my units pixel-perfect
and to always stay 1 pixel out of range of the enemies units or by
abusing timings ... like dropping a tank from a dropship and picking it
up after it attacked and before the enemy units/structures could react
and I see nothing of that with Deepmind's agent.
I was kind of expecting it to demolish players by making superior
decissions and by anticipating what the human would do on an extreme
level and I got disappointed. It does not yet play very smart ... and
after the initial low-resources gameplay which is pretty much all about
micro-management of your units it really should not have to rely so much
on said micro-management but on superior decision making. And that made
me think: How much room for improvement over a human world champion
player is there actually in Chess, Go and SC2? Can we measure how much
better a chess computer, Alpha Go or this SC2 AI agent actually is? How
much computing power would it take in order to compute the optimal thing
to do in an incomplete information game as complex as SC2? How much
better can any AI player get without relying on mere speed of execution
and multitasking but by relying on pure decission making? Also, how
would that gameplay look like? How can we say that rushing up a narrow
chokepoint was a bad decission if it worked out? Did the AI anticipate
that? Was it close to losing? I have no answers ... but clearly this is
not AIXI playing ...
Am 27/01/2019 um 00:53 schrieb Shashank Yadav:
This is significant though in case anyone missed it on Reddit here is
/a Starcraft fan's /
<https://www.reddit.com/r/MachineLearning/comments/ak3v4i/d_an_analysis_on_how_alphastars_superhuman_speed/>/analysis./
<https://www.reddit.com/r/MachineLearning/comments/ak3v4i/d_an_analysis_on_how_alphastars_superhuman_speed/>
Will be interesting to know how people on this list view this.
-
regards,
*The task is not impossible.* <https://muskdeer.blogspot.com/>
---- On Sat, 26 Jan 2019 18:00:46 +0530 *justcamel
<[email protected] <mailto:[email protected]>>* wrote ----
Deepmind just beat a top level player in a game many times more
complex
than GO.
https://www.youtube.com/watch?v=cUTMhmVh1qs
I really hope that AGI is near ... and thus the end of our
socioeconomic
insanity ...
*Artificial General Intelligence List
<https://agi.topicbox.com/latest>* / AGI / see discussions
<https://agi.topicbox.com/groups/agi> + participants
<https://agi.topicbox.com/groups/agi/members> + delivery options
<https://agi.topicbox.com/groups/agi/subscription> Permalink
<https://agi.topicbox.com/groups/agi/Tc6d84459c1e6c012-M00b4965731d0679a8a2a9923>
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/Tc6d84459c1e6c012-M768b68d4874b36bfa3a02b08
Delivery options: https://agi.topicbox.com/groups/agi/subscription
