<p>It isn't really a specific algorithm or something. It is a simulation of bots that live in a virtual environment. They can see, hear, move, speak, fire a laser weapon(computations for hands gets out of hand), etc. I have a prediction recurrent neural network that generates a tree of actions, and a situation analysis network(uses a genetic algorithm) to evaluate the different nodes. Bots have memory(can see past events), learn to predict the consequences of their actions.</p>
<p>One of the major goals of the project is to see the formation of weights in the prediction neural network(suggesting a more precise architecture to retry the experiment with). Also, we watch reruns of the bots lives(at points during which we predict interesting things will happen, such as the first encounter between isolated species), observing any signs of intelligence(working on the definition of that now). Finally, I devoted a lot of the paper(read 3-4 pages) too doing runtime analysis, and displaying the various obstacles the connectionist view holds. </p>
<p>A major issue is the combinatorial explosion of the action tree. So we implemented a sorting algorithm at each stage.</p>
<p>The thing I am most happy about is the very thorough runtime analysis.</p>
<p>I want to send you my paper(a PDF of it), but I'm not sure what the scientific rules are(aren't you not suppose to share until published?). Can anyone help me with this?</p>
<p>The project doesn't discover a new breast cancer protein(1st place team was from my school, drooling over plaque and cheque display in reception lobby as I write this), but I think it does propose and advance some fundamental ideas for the field. </p>
<p>How hard is it to get to semi-finalist for team? </p>
<p>Oh, there is a big emphasis on how human intelligence(the ultimate goal) is too complex for a simple hand crafted algorithm, and that the algorithm(very complex most likely) will be evolved. I also emphasized how now is the time to propose and test ideas, preparing for the major advancements in processing power to help carry out those experiments.</p>
<p>One of the biggest problems is that I am dealing with 5-6-dimensional optimization, with an unknown funcition(well I know for 3-4, which I worked out), but time is one of them, so you never know if you ran the program another day, it might have produced something.</p>
<p>I think AI(read aiming for general intelligence, not a specific vision algorithm) is a very trick field, but the benefits are enormous.</p>
<p>On that note, I have done a pretty cool project where one could simply point at the computer monitor, and the mouse would move to where the finger pointed(used a webcam).</p>
<p>Thanks for all the replies everybody, and sorry for derailing this thread. :)</p>