Version 0.4.4 release

[Brandon Rohrer]

BECCA version 0.4.4 is out this week. Here's what's new:

• Inputs are no longer constrained to fall between 0 and 1. Their distribution is learned.

• Reward is no longer constrained to fall between 0 and 1.

• A simpler similarity metric has been implemented, based on Manhattan distance, rather than angle. It’s faster to compute and appears to be equally effective.

• The State module and its associated class have been removed. 2D numpy arrays are sufficient to represent state. This sped up operation noticeably.

• The ability of each feature to accurately predict reward across the model has been factored into salience calculation. Performance on the grid_1D_noise world is now as high as expected.

• Deliberation in the planner was re-worked.

• Deliberate actions are taken at regular intervals, with observation occurring in between.

Soon you'll be able to download it from Matt's GitHub site:

https://github.com/matt2000/becca 

and from 

http://www.openbecca.org

If you're really in a hurry, you can pull it down from www.sandia.gov/rohrer

Brandon

[Matt Chapman]

It's posted on the website:

http://openbecca.org/project/core

I have a question about "Inputs are no longer constrained to fall between 0 and 1. Their distribution is learned."

If I understood correctly, you once told me that performance was significantly better for problems that could be modeled with binary input values instead of continuous values. Is that still true?

--
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]

Thanks Matt! 

And yes it's still true that each sensor input is treated as a graded binary input. The difference is that now if you have, for instance, a pixel value that varies between 0 and 255 you don't have to modify it to vary between 0 and 1 first. You can just feed it's raw value into BECCA and its distribution will be learned. BECCA will still interpret it as a binary on, off, or somewhere in between, rather than a continuous range. If you have a sensor, such as a position sensor, that you want to resolve at a number of different positions, you still need to bin your sensor into a larger number of on-off sensors in your world before passing them the BECCA.

Brandon

Brandon