Definitely a lot of choices. With Unity, I was able to run 50 players (100
eyes capturing 60x40 pixels) in a simple world at 10fps on my MacBook.
Within AngryBots, just two players at 10fps. Here are some more data points:
// Performance in world with one plane and a light.
/* PLAYER_COUNT | FPS
* 8 | 150
* 18 | 50
* 25 | 30
* 32 | 18
* 50 | 9
*/
To reduce the thousands of pixels streaming into NuPIC, I'm trying a
Convolutional Neural Net with Theano. Will post with how it goes.
I imagine that 2D and especially 3D physics engine provides interaction
> subtleties that would test AI unlike any other method. Very slight
> movements can result in drastic changes to the world and the agent's body.
> In other words, the degree of freedom in the control signal space can be
> arbitrarily large.
Totally. Hoping that 6 degrees of freedom (move_x, move_y, look_x, look_y,
jump, shoot) will yield insights into sensorimotor networks that can be
extended further.
Critterding looks cool, like a Polyworld with a richer environment. This is
actually something that the creator of Polyworld mentioned was a
constraining factor in the sophistication of the neural nets that emerged
from in the world based on 90's tech stack. I'd like to compare with the
neural nets that evolve from within Critterding in the much richer
environment.
On Mon, Mar 17, 2014 at 6:56 AM, SeH <[email protected]> wrote:
> Gazebo / ROS is a more complete solution, whether in desktop or web
> mode. SpacegraphJS is still just a prototype. Its Ammo.JS + Three.JS
> browser-based design does have significant performance limitations but
> could be optimized (ex: WebCL)
>
> The browser-based front-end can be responsible for initiating and
> controlling the interaction with the cognition engine, probably through
> websockets for lower-latency I/O. The attached cognition engine back-end,
> serving a websockets connection, would be entirely decoupled from the
> front-end, receiving sensor data and transmitting motor commands. RLGlue
> follows a similar model: http://glue.rl-community.org/wiki/Main_Page
>
> Craig's AngryBotsInAiWorld <https://github.com/aiworld/AngryBotsInAiWorld>is
> worth further development too but possibly without a server involved.
> Have you seen Quake and other FPS ported to WebGL? Those older FPS do not
> involve a physics engine so they are less computationally intensive and so
> can more easily support larger worlds and multiplayer action. Bots in such
> a world might overwhelm current versions of NuPIC if their entire "retina"
> of thousands of pixels were used as sensory input. Instead maybe they can,
> at first, distill the perceived scene to numeric geometry (without
> cheating) such as knowledge about distance to nearby objects like walls and
> players, measured at discrete intervals (ex: every 10 degrees rotating
> around the bot). This is how conventional FPS bots must operate.
>
> I imagine that 2D and especially 3D physics engine provides interaction
> subtleties that would test AI unlike any other method. Very slight
> movements can result in drastic changes to the world and the agent's body.
> In other words, the degree of freedom in the control signal space can be
> arbitrarily large.
>
> For high performance simulation, a native C++ OpenGL game engine is still
> the best. SpacegraphJS is a JS port of SpacegraphC
> https://github.com/automenta/spacegraphc2 which was originally inspired
> by Critterding http://critterding.sf.net which could all be connected to
> a cognition engine in the ways described here. There's also several
> open-source Minecraft clones.
>
> So, there's many possibilities. Difficult to choose. Any ideas?
>
>
> On Mon, Mar 17, 2014 at 12:27 AM, Craig Quiter <[email protected]> wrote:
>
>> Wow, very cool to see another 3D platform for testing
>> sensorimotor algorithms, all in the browser no less! I had looked into ROS
>> / Gazebo which can also render via three.js, but this still needs several
>> supporting services on Linux.
>> http://gazebosim.org/wiki/Tutorials/CloudSim/IntroductionToiPyNotebook
>> Awesome to see sensors, rendering, and physics all running in the browser.
>> Great work!
>>
>> On Sunday, March 16, 2014, Chetan Surpur <[email protected]> wrote:
>>
>>> This is awesome! Are you building this?
>>>
>>>
>>> On Sun, Mar 16, 2014 at 9:58 AM, SeH <[email protected]> wrote:
>>>
>>>> In the Sensory-Motor integration talk, Jeff mentions testing in a maze
>>>> with a few sensors and motor controls. Here is a maze world added to this
>>>> 3d
>>>> physics simulation <https://github.com/automenta/spacegraphjs>.
>>>>
>>>> The bot is a simple worm with angular motor controls for each segment.
>>>> It "sees" through raytracing a number of vision paths to find the distance
>>>> to the nearest collided object (which could also provide the observed color
>>>> as an R,G,B vector). The current bot controller is operating entirely
>>>> randomly but this could be replaced with NuPIC, OpenBECCA, or some other
>>>> cognition engine or reinforcement learning algorithm.
>>>>
>>>> Any type of world or bot geometry, motors, and sensors could be
>>>> constructed from the 3d primitives and controllable constraints / joints in
>>>> the physics engine API - in this case, Ammo.JS (which is an automatic
>>>> Javascript port of the C++ Bullet 3D physics engine).
>>>>
>>>>
>>>> [image: Inline image 1][image: Inline image 2][image: Inline image 3]
>>>>
>>>>
>>>>
>>>> On Fri, Feb 28, 2014 at 4:36 PM, Matthew Taylor <[email protected]>wrote:
>>>>
>>>>> http://www.meetup.com/numenta/events/168671932/
>>>>>
>>>>> This event is in San Jose on March 14. There are only 50 seats, so
>>>>> RSVP quickly!
>>>>>
>>>>> ---------
>>>>> Matt Taylor
>>>>> OS Community Flag-Bearer
>>>>> Numenta
>>>>>
>>>>> _______________________________________________
>>>>> nupic mailing list
>>>>> [email protected]
>>>>> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> nupic mailing list
>>>> [email protected]
>>>> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>>>>
>>>>
>>>
>> _______________________________________________
>> nupic mailing list
>> [email protected]
>> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>>
>>
>
> _______________________________________________
> nupic mailing list
> [email protected]
> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
