My little Siemmons Westinghouse Question

Colleges and Universities A-Z Massachusetts Institute of Technology
1

<p>Yeah. I could have posted on the other forums, but thought that there are a lot more "winners here."</p>

<p>I am a sophomore, with a few unique ideas/projects I have been working on for a few years. Here is a description of one of them:</p>

<p>Imagine a digital 3D world, where we have little triangle shaped species. Each Triangle recieves an image of what it is seeing out of it's two eyes, as well as the sounds it heres, what it is feeling(if it hits something). It can speak, push objects, and shoot a laser(someway to hurt each other). Basically, I am going to do a mass simulation of hundreds of triangles interacting with each other, using multiple massive unsupervised back-propagation neural networks(possibly fiddle around with recurrent nets...), a set of instincts such as pain to amplify the signal, and genetic refining, so that only the best survive. I will probably implement them one at a time. My goal is to see if the triangles understand what they are doing, or if they show any signs of intellegence. The neural networks will probably have thousands or inputs, and hundreds of hidden layers. I will probably code it in C++ or C#, though I am leaning towards C# because speed is not really of concern(using c++ dll's for the number crunching of the error training.) because it has nice access to the internet, as well as managed directx, which is awesome.</p>

<p>Also, since the project has some many computations, it is too slow for any realtime. I basically calculate everything overnight, saving data in a little file.
I designed a program that can read the file, and display what is happening like a movie, cept you can move around.</p>

<p>I heard of the competition a few months ago, and though it would be fun to apply, and see what other people are doing.</p>

<p>I also have some other projects I am working on.</p>

<p>A project called the Research Project, which attempts to search the internet, and put together some kind of coherent report on the subject. Basically copy pasting valid stuff from websites. I am working on the algorithm for both ensuring validity of info through multiple sites, and picking up other keywords for broader search.</p>

<p>In the meantime I am building a Chess Program which I placed a wager on to beat my friend, who is 1st in state for chess.</p>

<p>I can't post much longer(I was experimenting on me and my brother by playing tapes at night, and then recording answers to questions in the morning) cuz my mom thinks i am making my bro late for school.</p>

<p>In short, the siemmons westinghouse people won't lemme enter without some other people. Should I just pick up some dummy at school, and claim we are a team or what? Also, is MIT the kind of school that I should aim for? I live in NC. SAT = 1480. Rank 1/400. Unsure about what to do.</p>

<p>PS: I am self studying calc bc this year, while taking Math Methods I IB at school. My GC says I shouldn't take it, and if I do, I should just take Methods II, so I could help others when they need me.</p>

2

<p>Very interesting and creative projects, although nothing which really will be original research. I would suggest repeating experiment with squares and circles. What's the difference? You could probably get some insight out of that - why did different species evolve to be different shapes?</p>

<p>Shoot for MIT, man - you are a creative person.</p>

3

<p>That is a good point. How exactly do I find out what is real research and what is not. Also, are the projects I mentioned a waste of time? Because if they are, than I see no point in spending hours reinventing the wheel, instead of trying to advance it. What exactly do you mean with Circles and Squares? Do you mean like different species? That would be interesting, but extremely ardous.</p>

4

<p>Forgot to mention: the whole point of that triangle project is whether unsupervised networks will provide any sort of intelligence. I chose that field, because most AI researchers are afraid to tackle it: getting results is much more difficult. </p>

<p>Thanks for the advice. I need people to help me with deciding the specifics of what to do. Now that I am at it, is it possible for this sort of project to win anything at Siemmons Westinghouse level? I mean, I am not getting help at RSI or something, so I am really on my own.</p>

<p>If it works, than that means that virtual showed signs of AI without any specific human programming.</p>

5

<p>I highly suggest getting in contact with researchers at local universities and trying to find a mentor, bringing up the ideas that you've had.</p>

6

<p>Oh, don't be discouraged by those people- you're brilliant for thinking of something like that as a sophmore. Although yes, neural networks and evolutionary algorithms aren't new, programming something like that makes you learn++ - a very good experience. Just keep thinking outside of the box, reading for fun until something comes to you. It's hard, but it will eventually.</p>

7

<p>Hey Sagar, I can be your dummy.... lol
along with Nirav :)</p>

<p>Then you can work on it and we won't mess anything up hehehe
Those projects are great, too bad I can't think of them myself.</p>

<p>Go ahead and study Calc BC, but wait a year to take the exam and take Methods II next year. Then take some college math course senior year.</p>

<p>Good luck on SAT's, c-ya monday</p>

8

<p>just wondering... to those that responded, are any of u semis or finalists this year in siemens? and how prestigious is it really for a semi?</p>

<p>and yes, highly recommend that you discuss the topic with a knowledgeable person in the field because there is almost no other way to find out if it is original... unless u urself have the equivalent expertise of a phd in the field</p>

9

<p>Well, my dad has a phd. However he says that neural networks aren't popular anymore because they don't show much promise.</p>

<p>Also, a q about Calc BC. The differentiation is really easy, but the integration is harder! I thought they were suppose to be even in difficulty.</p>

10

<p>semi for siemens isn't much, but finalist will generally "get you in"</p>

<p>and sagar, integration is boatloads more difficult than differentiation. it's almost like comparing arithmetic and geometry</p>

11

<p>Can the project I described above actually win me anything in siemmons westinghouse?</p>

12

<p>Or at least something like semi-finalist status?</p>

13

<p>Bumpity Bump</p>

14

<p>Your projects are very creative, yet lacking in applicability to the real world. It would be "safer" (aka more likely to win a prize) to tackle a more popular subject. Then again, I also recommend NOT doing seimens because you want to win. I worked on a research project and even though I got nothing from seimens, I still had a great experience. Also, Intel is about 10000x more prestegious than seimens, so if you're aiming high, try that one.</p>

15

<p>Sagar - </p>

<p>Neural networks are NOT ment to have HUNDREDS of layers. Any introductory text will tell you that anything more than 4 layers becomes tremendously complex. Also for your needs, 3 layers should be more than enough. It would be wiser if you make ur individual "triangles" entities of their own and place many of them in an environment. You are overcomplicating things. </p>

<p>A much simpler approach would obviate the need for neural networks. They are designed to adapt to an environment and perform a specific tast VERY well. It would be better if u looked at genetic algorithims instead.</p>

16

<p>Yeah, I heard about the whole 3-layer industry standard. But here is my thinking: if a neural network can represent millions of if then and switch statements, then theoretically, if the network only need 3 hidden layers, then the other would layers would act redundant. But Since My input data is huge, and... hard to explain, but here is my plan:</p>

<p>train a network that is part of the triangles brain, to take inputs that include the last x states of the world input, and output from main brain, to predict what will happen. Then, train another neural network to, given a predicted output, and a database full of instincts and goals, which themselves were determined by the triangle, to determine how to modify the output of the triangle's main ANN. g2g, ttyl, class has ended...</p>

17

<p>That seems a bit difficult. If you want to achieve anything similar to AI, as i think you are, genetic algorithims are the way to go. </p>

<p>Im more familiar with the hardware-based ANNs, but im pretty sure if u make neurons with only logic operators (w/o weighted inputs, summers and transfer functions), its just a glorified von numann system. Hope this helps.</p>

18

<p>Yeah, I was gonna have genetic algorithms, but I think I need to use a combination of ANN's, Gen Algrth's, and that other thing, fell of my tongue now, but it based on whether the organism gets hurt. </p>

<p>Also, if I have an infinite training set, then why can't I have more hidden layers? wouldn't it simply make the ANN more stable?</p>

19

<p>ok what exactly do you mean by an "infinite training set?"</p>

<p>the 3-4 layer ind. standard is in place because the operation of the system becomes wayy to complex. Im working on a ANN robot for (hopefully) ISTF and it is a pain in da butt. Trying to keep it at 3 layers. </p>

<p>And to everyone out there - dont bother with siemens if ure from jersey like me. They hate jersey people for some reason.</p>

20

<p>ok. lets suppose that we are working in a virtual world. We could create our own little scene dynamically, and train the network. Think! You could build a raytracer to have super-detailed scenes. You could then ask the ANN to determine all these aspects of the scene, etc. Because you can dynamically create scenes, you can have the computer make scenes for you, so that you could have as much training as you want.</p>