https://www.aicrowd.com/challenges/unity-obstacle-tower-challenge

As a starter, they provide:

https://github.com/google/dopamine

Which is more or less a test driver that loads little tensorFlow scripts...

Because this is my computer we are talking about, TensorFlow didn't
install right, a few of its dependencies wouldn't build. -- cloud
computing is available from GooG and probably amazon too, don't wanna go
that route, easy debugging and all...

My machine here has 32gb ram, a GTX980Ti, Gentoo Linux. AMD R7-1800x,
Motherboard is a fairly entry level desktop model, only one GPU slot,
micro-ATX, [want a threadripper. =|   ] , will cost $2,000 to upgrade
this machine, and more if I want to upgrade the video card too.


Ok, lets look at the problem.

It is NOT an AGI complete problem, it does not contain unbounded
emergant concepts, nor does it require communication or cooperation.

However, as the whitepaper explains, it provides the next challenge
after basic 8-bit (more like 4-bit) Atari games. It provides deeper
perceptual challenges, an agent will have to be able to separete the
basic play-relevant form of the room from any of a number of visual
styles, as well as realistic lighting situations such as shadows, near
darkness, light rays from windows, etc. It must also deal with moving
objects in the environment and a variety of specific puzzles, such as
moving a block to a specific spot on the floor while avoiding another spot.

The paper reported that existing techniques got stopped by problems
requiring the ki. There are 3 representations of the ki, the 3D
representation in world, the screen icon version representing a ki in
inventory, and the visually distinct version used to label the doors
which require it.

I don't think conventional DL techniques are going to cut it. There
needs to be a more sophisticated system to handle the de-coupling of the
play-relevant form of the room from it's visual styling. In number
theory terms, it needs a way to factor the input like a number, discard
the irrelevant factors, and base decision making on what remains.

I'm a bit too video game adicted to get super serious about this but
I've been putting together some notes while I work to install and
understand the software I need.

[several days have passed since I started this draft, there's more to
say, but going to post this portion now...]

-- 
Please report bounces from this address to a...@numentics.com

Powers are not rights.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T82aac836327740b0-M135911d7391e4ba27a31117d
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to