Contest?

[Matt Chapman]

I curious if we have lurkers watching this list just in need of a small enticement to get involved.

Would anyone be interested in a contest with small prize (Amazon gift card just large enough to buy a couple books or something) to create a World to solve a particular competitive task?

If something like that would inspire you to spend an afternoon hacking on BECCA, say yay.

For those who have yet to try out BECCA, yes, you can in fact make something interesting in an afternoon.

All the Best,

Matt Chapman
Ninjitsu Web Development
http://www.NinjitsuWeb.com
ph: 818-660-6465 (818-660-NINJA)
fx: 888-702-3095

http://www.linkedin.com/profile/view?id=13333058

If you want to endorse my skills on Linked In, the most valuable endorsements to me are "Open Source" and "Software Development."

--
The contents of this message should be assumed to be Confidential, and may not be disclosed without permission of the sender.

[Brandon Rohrer]

Matt, that is a super cool idea. I would be pleased to dedicate the BECCA Foundation's Endowment to fund the prize. Right now the balance stands at $23.19

Brandon

[Matt Chapman]

Cool. Ninjitsu Technologies will match that and round it up to an even $50, if we get more than one new contributor interested in participating. If we get 8 or more, I'll come up with a second prize too. Are there even 8 members of this list? ;-)

--

[SeH]

i'd like to see a BECCA vs. Q-Learning challenge, not sure which is better yet or if they are actually simialr algorithm underneath

[Matt Chapman]

I was thinking more along the lines of a BECCA developer vs BECCA developer challenge. I.e., who can develop the best World/task for a given goal, like learning to win at a simple game like Connect 4. Each BECCA would be pitted against a narrow AI that uses a traditional tree-search algorithm to play, and would be scored based on how quickly it learns to tie or win by observing it's opponent.

To answer your question, I believe BECCA can be cast as a kind of extension to Q-Learning. I'm sure thats an oversimplification at best, and Brandon can correct me if I'm wrong.

[SeH]

there should be a few software kits for doing reinforcement learning experiments in which demos could support becca, q-learning, and other exotic combinations of software agents

i would eventually like to see agents operating a physics simulation like http://critterding.sf.net or www.rigsofrods.com both are open source engines unlike Unity3D which is currently proprietary

but at least finally unity is going to linux: http://blogs.unity3d.com/2012/07/03/linux-publishing-preview-what-how-and-wherefore-3/ though i would like to see opensource agents on an opensource simulator all the way through so that simulation and simulant can be integrated

[Azriel Fasten]

Lurker here, would be interested. I'd consider using Panda3D, since it is written in python. I've been watching and wanting to hack on becca for a while.

--

AYF

[Matt Chapman]

That's interesting, and suggests another possible angle for the contest:

Create a BECCA agent for the open-source virtual world of your choice, and Brandon and/or I could judge them on criteria such as:

- Best demonstration of BECCA's strengths
- Best reproducible AGI Experiment
- Best third-party software integration
- Best improvement to BECCA core
- etc...

[Alejandro Dubrovsky]

My first idea with becca was to try it with MAME games. I've got an
interface to control MAME games (eg galaxian) from python at
http://organicrobot.com/mame/ . The main problem is that I get easily
distracted.

Presenting the whole screen as input would very likely be too many
inputs, but feature extraction from the old games is easy since objects
in the game tend to be a short loop of sprites so matching an exact
pixel arrangement would be successful at picking out objects. I will
eventually get around to it, but if anyone feels like trying it out in
the meantime I can provide a lot of the supporting code.

> fx: 888-702-3095 <tel:888-702-3095>
>
> http://www.linkedin.com/profile/view?id=13333058
> <http://www.linkedin.com/in/chapmanmedia>

>
> If you want to endorse my skills on Linked In, the most
> valuable endorsements to me are "Open Source" and
> "Software Development."
>
> --
> The contents of this message should be assumed to be
> Confidential, and may not be disclosed without
> permission of the sender.
>
>
>
>
>
> --
> All the Best,
>
> Matt Chapman
> Ninjitsu Web Development
> http://www.NinjitsuWeb.com
> ph: 818-660-6465 (818-660-NINJA)

> fx: 888-702-3095 <tel:888-702-3095>
>
> http://www.linkedin.com/profile/view?id=13333058
> <http://www.linkedin.com/in/chapmanmedia>

>
> If you want to endorse my skills on Linked In, the most valuable
> endorsements to me are "Open Source" and "Software Development."
>
> --
> The contents of this message should be assumed to be
> Confidential, and may not be disclosed without permission of the
> sender.
>
>
>
>
>
> --
> All the Best,
>
> Matt Chapman
> Ninjitsu Web Development
> http://www.NinjitsuWeb.com
> ph: 818-660-6465 (818-660-NINJA)
> fx: 888-702-3095
>
> http://www.linkedin.com/profile/view?id=13333058

> <http://www.linkedin.com/in/chapmanmedia>

[Brandon Rohrer]

I love where you guys are going with this. MAME looks like a great place to apply BECCA to some more interesting tasks, and Panda3D looks like a super cool way to create more games/tasks. SeH, actually Nick Malone is doing some work right now to get Q-learning in the right format to perform the benchmark tasks. We've learned that all but the simplest tasks result in a state space that is far too large for Q-learning to enumerate, so it doesn't even get started learning on them. Hopefully before too long we'll have values for the grid_1D set of tasks though.

What I'm seeing here is that our interests are diverse and ambitious. Please jump in (with all your spare time) and please keep us all posted! I'm excited to see what you create. I'll try to support by fielding BECCA technical questions and adding features as I'm able. Matt I'll leave it to you to make the rules for and organize any formal competition, but I'll support it however I can.

Brandon

[SeH]

create more games/tasks. SeH, actually Nick Malone is doing some work right now to get Q-learning in the right format to perform the benchmark tasks. We've learned that all but the simplest tasks result in a state space that is far too large for Q-learning to enumerate, so it doesn't even get started learning on them. Hopefully before too long we'll have values for the grid_1D set of tasks though.

great.. which implementation of qlearning?  here's links to earlier work in critterdroid that provides a common interface for BECCA and Q-Learning:

https://github.com/automenta/critterdroid/blob/master/src/com/critterdroid/bio/brain/QLearningBrain.java

https://groups.google.com/forum/?fromgroups=#!topic/becca_users/ghv2xA4BMwM

iirc the Java QL implementation used in critterdroid performed much better than the python version of BECCA via the JVM/Python TCP bridge

[Brandon Rohrer]

I haven't taken the time to step through the details, but when I do I'll pass on what I find out. I also regret not spending time yet with your critterdroid interface. It is super cool that you've put that together. I've been focusing on getting the basics working right. BECCA still trips over its own feet in a lot of simple instances, and I'm ironing those wrinkles out. But I'm really excited to extend it to more interesting environments and tasks, especially ones based in the physical world.