Tuesday, 1 April 2008


I have been thinking about what to do next. So far, I have been following up on one question after another and never really solidifying where I have been. Part of it has been due to my poor research skills, a better approach would be to write and present my thoughts as I go instead of keeping it all in my head where, like a house of cards, it gets increasingly difficult to keep up.

So this is what I am thinking, a few of the things that would be a good idea to do would be to sketch out schedule, polish my literature survey, skim through my PhD guide and look into implementation details such as available tools.

A little while back, I did look into a few tools but nothing really came out of it. I suspect that papers on evolutionary emergent communication such as Cangelosi (2001) as well as Cangelosi and Harnad (2002) used tools such as MATLAB. I also suspect that other papers on embodied artificial intelligence used their own tools.

The following address is to research done by A. Cangelosi.


Cangelosi et al. (2007) used the Open Dynamics Engine (ODE) to simulate physics. Right now, I am more interested in what tool Cangelosi and other researchers to simulate neural networks.

My quick search for neural network simulators returned a number of candidates. I do prefer neural network libraries that are compatible with game engines.

An ONLAMP page at the following address contains a few ideas about neural networks for games. The ideas are from the authors of "AI for Game Developers," D. M. Bourg and G. Seemann.


The super cool thing is that the book is also available at Google Books. Having this much awesome information available to me makes me feel spoilt and a tad bit evil.

Please note that I feel that neural networks need substantial additional augmentation in order to be truly useful.

Unfortunately, "AI for Game Developers" on Google Books lacks a large number of relevant pages. I will need to get the book to see how they implemented their neural networks and with what libraries.

Okay, I was looking at the provided example code for the book and saw that the authors wrote their own neural network class. While this is useful, I would prefer a more sophisticated solution such as MATLAB because of features that help analysis such as simply plotting graphs. This may be a bit too much to ask.

I did hear that MATLAB might be able to export code for a neural network after it has received specifications. This code can then link with game engine code.

I compiled and ran the example neural network program. I am not sure of what I was seeing because I have not read the related chapter. I am not sure of how much further to follow this track.

Other ideas include using Octave, Python and FANN. I do like the look of FANN because it is free, open source, portable and fast.

I think it would be a good idea to revisit the other software I looked at earlier and then quickly choose one to prototype some ideas.

No comments:

Post a Comment