[axcho]

I've been interested in playing with some Artificial Life projects in Second Life, which is a virtual world where you have an avatar and make buildings and stuff. I haven't actually done anything yet, but reading this book made me want to try making virtual pets that use HTM.

I think that Second Life would be a good place to do this because it provides its own virtual environment for the creatures to live in and interact with. The nice thing is that since there is no need for low-level vision processing, the creatures might be able to be fairly intelligent with fewer neurons, since they can already see the world in terms of objects. (and language-processing would be awesome if it's possible )

Another reason is that if you sell them as virtual pets (something like the NN-based creature in the game Black & White), not only can you make money, but people would take the time to train them in the necessary way, like a pet or baby. Recently I heard a story on the radio about some scientists who found that Bonobos could learn to understand human language if they were taught like human toddlers are. So selling virtual pets might be the easiest way to get a lot of AI brains trained effectively.

Anyway, I think this would be a fun way to test out the theory and maybe get famous and make some money at the same time. What do you think of the idea?

[FreeSynapse]

I think this raises a bigger question of how and whether HTMs could engage in reinforcement learning, which is different from supervised learning.

Also, depending upon the license under which HTM is used, one may or may not be able to continue to train the HTM after it has been "sold". For instance, as I understand, for licensing reasons, you cannot use the MATLAB neural net toolbox to create an independent program that by itself trains neural nets.

[Numenta-Phil]

Continued learning is one of the issues that I think HTM has to address. There are some applications, embedded in automobiles or phones, maybe, where the manufacturer doesn't want different devices to behave differently based on the environment they've been exposed to.

However, for the most part, I think that Numenta will be providing you with the ability to keep your application learning as long as you want it to (obviously, training is more compute/memory intensive than inference or prediction).

As for the virtual pet idea, I think it's great! An interesting project to work on. Plus, read the book "Play Money" (not sure of the author), who discusses that lucrative online 'game' market for third party virtual goods.
_________________
Phillip B. Shoemaker
Director, Developer Services
Numenta, Inc.

[axcho]

Hmmm, it seems like no one knows how to combine this model of intelligence with reinforcement learning. Oh well, if I think of a possible solution I'll post it here.

Also, when I said HTM, I was just using that as an abbreviation of the general memory-prediction model presented in the book. I'm not exactly about to use MATLAB to program a Second Life script...

[FreeSynapse]

[axcho]

[Stack/Heap Error]

You guys should check out these two cats I found in Second Life. I don't know what kind of programming is in them, but they were maneuvering around in the virtual world on their own. I've visited them a few times now where they reside, up on some mountaintop platform. I fed one of them just today, because when I clicked on it, it said it was hungry.

Kind of simplistic maybe and surely not a true animal level intelligence, but seemed like these cats had some kind of basic AI or at least give that illusion well?

This is the place, seems its all centered around creating animals:

http://world.secondlife.com/place/d1677b9f-8b80-2cdb-16f2-9ade252c7e0c