Please advise me on this project!
Author
Hi guys. I haven't been on the AI forums for a long time(school, etc), but I did want to come back and get some advice from you guys.
I've worked on a small research project, which I wrote a paper for and entered in the Siemens Westinghouse Competition. I didn't place, but I am now preparing for the Intel Science+Engineering Fair.
I was wondering if you guys could look at my abstract, and better yet my paper, and tell me what you think. I am not sure if I am heading in the right direction.
This can be found here.
Author
Nobody?!?
Here's the Abstract:
Abstract
This paper discusses a novel approach to the development and evolution of connections that form between neurons in the brain and the emergence of intelligence, as well as the construction of a simulation designed to test these ideas, with a heavy emphasis on the runtime analysis of the simulation. The concepts and methods proposed in this paper contribute to the disciplines of artificial intelligence and neuroscience, with strong ties to the cognitive development of the human infant.
The simulation consists of a virtual species that learns to adapt to a hostile, dynamic and evolving environment, with members of the species differentiated only by their intellectual abilities. The brain of a member of this species is composed of two neural networks: the prediction network, a recurrent neural network that employs a derivative of the standard back propagation algorithm, and a situation analysis, a feed forward neural network trained by a periodically run genetic algorithm, enabling the evolution of the species. Together, the two networks generate a large look-ahead tree, and evaluate it.
The most recent results of this simulation support the runtime analysis presented, suggesting the strong dependencies of the discipline of artificial intelligence on the hardware advances in coming years.
Here's the Abstract:
Abstract
This paper discusses a novel approach to the development and evolution of connections that form between neurons in the brain and the emergence of intelligence, as well as the construction of a simulation designed to test these ideas, with a heavy emphasis on the runtime analysis of the simulation. The concepts and methods proposed in this paper contribute to the disciplines of artificial intelligence and neuroscience, with strong ties to the cognitive development of the human infant.
The simulation consists of a virtual species that learns to adapt to a hostile, dynamic and evolving environment, with members of the species differentiated only by their intellectual abilities. The brain of a member of this species is composed of two neural networks: the prediction network, a recurrent neural network that employs a derivative of the standard back propagation algorithm, and a situation analysis, a feed forward neural network trained by a periodically run genetic algorithm, enabling the evolution of the species. Together, the two networks generate a large look-ahead tree, and evaluate it.
The most recent results of this simulation support the runtime analysis presented, suggesting the strong dependencies of the discipline of artificial intelligence on the hardware advances in coming years.
From your website I gathered that you were still in high school, and taking that into account the project seems very advanced. Definitely way more advanced than anything I was doing in high school. I can imagine you jumping right into research in college and doing some exciting stuff.
From reading the abstract and skimming the paper, some of the language and claims are maybe a little too vague and grandiose for an AI conference or journal. That is, instead of saying that your work contributes to AI and neuroscience and talking about the fields in general, I would say what technically and specifically your work contributes and how this is different from previous work. For example, emphasizing your two neural network approach and comparing it to single neural network approaches to similar problems. The second paragraph of the abstract is right on. You might consider looking at previous papers from a conference or journal and modeling your paper's format on them. AAMAS might be a good conference to look at for your work.
Another thing, if you were submitting this to an AI conference or journal, depending upon where you were submitting it you might consider changing some of the language and de-emphasizing the biological slant. That is, instead of saying "member of the species" and "brain", say "agent" and "control program", and talk less about neuroscience if your work is just using neural networks.
From reading the abstract and skimming the paper, some of the language and claims are maybe a little too vague and grandiose for an AI conference or journal. That is, instead of saying that your work contributes to AI and neuroscience and talking about the fields in general, I would say what technically and specifically your work contributes and how this is different from previous work. For example, emphasizing your two neural network approach and comparing it to single neural network approaches to similar problems. The second paragraph of the abstract is right on. You might consider looking at previous papers from a conference or journal and modeling your paper's format on them. AAMAS might be a good conference to look at for your work.
Another thing, if you were submitting this to an AI conference or journal, depending upon where you were submitting it you might consider changing some of the language and de-emphasizing the biological slant. That is, instead of saying "member of the species" and "brain", say "agent" and "control program", and talk less about neuroscience if your work is just using neural networks.
Author
Quote:Original post by Anonymous Poster
I believe he is in MIT, judging by his blog, not High School.
LOLOLOLOL... ok i'm done.
I WISH I WAS IN MIT!!! That's where I dream of going. I'm currently a junior in High School.
I'm contacting a professor at Duke University to take a look at my project. I did this all by myself, so it isn't the best it could me, and I'm really tired of continuously running into dead ends, so I guess I really just want like an older mentor who I can bounce ideas off of. Working alone is... lonely :(.
By the way, if you guys could read the paper on my website. I modeled it after some CS papers I found online.
The direct link to the paper is here.
Thanks for the responses!
I thought it showed a lot of knowledge, especially for your age. At your rate of learning you will know about 10 times more than me in a few years.
There are some minor criticisms that I think will help for future projects.
Your project was far too ambitious. That isn't a real problem as it does show you are ambitious. But what you were planning to do was immensely ambitious beyond words. I think you can probably see that your project could be split into many seperate parts, each part being a project in itself. But I did find the principle of what you did interesting.
I did occasionally find some of the language you used in the paper a bit over dramatic, for example the "and a new age shall dawn" phrase at the end. But then it does show you have vision so I don't think it's a problem.
There are some minor criticisms that I think will help for future projects.
Your project was far too ambitious. That isn't a real problem as it does show you are ambitious. But what you were planning to do was immensely ambitious beyond words. I think you can probably see that your project could be split into many seperate parts, each part being a project in itself. But I did find the principle of what you did interesting.
I did occasionally find some of the language you used in the paper a bit over dramatic, for example the "and a new age shall dawn" phrase at the end. But then it does show you have vision so I don't think it's a problem.
Author
Quote:Original post by RunningInt
I thought it showed a lot of knowledge, especially for your age. At your rate of learning you will know about 10 times more than me in a few years.
There are some minor criticisms that I think will help for future projects.
Your project was far too ambitious. That isn't a real problem as it does show you are ambitious. But what you were planning to do was immensely ambitious beyond words. I think you can probably see that your project could be split into many seperate parts, each part being a project in itself. But I did find the principle of what you did interesting.
I did occasionally find some of the language you used in the paper a bit over dramatic, for example the "and a new age shall dawn" phrase at the end. But then it does show you have vision so I don't think it's a problem.
Yes well... when you've get 3hrs of sleep per day, and its the night before submission... people get dramatic!!! lol. I did the project for about a year on and off, sometimes putting in 4 hours a day, sometimes just thinking stuff out in my mind. The paper took about a month, on and off. Usually, I wrote it from 12am-3am.
It was very stressful, especially since I entered Siemens as a team(I did project independently, but picked up a programmer dude who is also a junior 3 weeks before submission, and he was my "teammate"), and as team leader, I had to fill out a lot of paperwork. I just constantly kept thinking, screw the project, what if we are rejected for some paper-work mistake!!!
The conflict I have is that even though these individual parts might not work independently, they would work when put together. So it's like I need to have everything work together.
The ideal situation would be that I'm not in HS, doing hw all the time(and trying to maintain a shred of a social life), and could devote myself to this full time.
Overall, it's excellent for high school work!! In fact, I'm very happy for you and surprised that you got all of this hooked up correctly in the first place (congrats!) and that you were able to come up with a plan for analyzing the results! I forsee you as being a great researcher in the field as well, as this is actually a great stepping stone toward some great work. It's clear that you spent a lot of time on it, and that you've also learned a lot from it as well. Many of the students entering that competition actually do their projects for a class or during a stay at a university, so I think this shows GREAT initiative on your part to do it during your free time -- I'm very very very happy for that since very few students will take that initiative if it's not given to them as classwork. If only you were able to understand the networks and the mathematics behind them better than at a very high-up superficial level then you would really have been able to go somewhere with this!
Sorry if I sound harshly critical below, but I'm comparing it to a paper that a researcher might write in the field (which is what I'm assuming they look for?). If I got something wrong (which is very possibly the case) then please let me know. The paper's largest flaws are the lack of a good introduction and the lack of a good results/discussions section, but read on for more on this. The anonymous poster above gave some great ideas -- heed his words! You state that your research is 'novel', but you never explain why. I am even left questioning why you even looked into this in the first place, and how it may be an important model for the future (use specifics).
In your intro, you mention that your model "retain the basic features that comprise human intelligence" – the field doesn't even know what these are, and frankly the ANNs you're using don't retain them all. (In fact, it would make for an interesting analysis to explain which learning features your model retains and at what gain in run-time compared to for instance Grossberg's Adaptive Resonance Theory model, which does a decent job at incorporating most learning features). <br><br>Saying things like "The field of AI has made limited progress over the past 50 years" is NOT a good idea, since it is absolutely not true! Computers barely existed then! You're basically saying that the entire field was worthless since its inception, so why are you going into it?<br><br>The end of the intro and section 1.1 should be removed completely – they are far too general. Something that REALLY sticks out is the absence of a good introduction. I don't mean to be overly critical, but your reference section is rather short and primarily consists of websites, and your introduction doesn't have a single citation. I bet that they are expecting something along the lines of a literature review in the introduction to illustrate that you've read the relevant literature in the field (academic papers) and that you know how to apply it to make better things. Further, what does your research do that scientists have not done in the past? How is your research unique? Why is it important? Your current intro doesn't even discuss why you chose the networks you are using or even their properties which enable them to learn at all! In general, putting websites in your reference section is not seen at all in the literature due to validity issues. If your references are published in a reputable journal, then you at least know that an expert in the field wrote it. Right now, I get the impression that you just came up with this idea out of thin air for no reason and you're just going to plug it all together haphazardly and hope you get some good results. At the end of the intro you should perhaps say you hypothesis for what you believe will happen (in precise terms) and why, then a short bit on how the results affirmed or disproved your hypothesis (this should also go in your abstract, but make it very brief – most of this will be discussed in the discussion/conclusion section).<br><br>It's starting to sound like a computer game now – WHY did you choose these particular characteristics and actions for the bot? You may want to remove the first paragraph in 2.2 entirely – scientists know all about seeding the weights randomly, it's not novel and YOU certainly didn't model it after observing babies. <br><br>In general, you should always have a caption for all of your figures and graphs as well as a figure/graph caption number (i.e. FIGURE 1. xxxx).<br><br>You're doing a lot of hand-waving in 2.3x with describing the networks in very vague terms. I'm not quite sure you have an understanding of what they actually do or how they actually work, and I'm left wondering why you chose these at all. Your page 13 figure is good (although captionless), but to me it begets the question of why you didn't use an alternate approach such as Markov Processes which predict the future solely based on the past x inputs. Why did you choose neural networks at all? (I'm not saying these would be a better choice, but you never address the issue.)<br><br>"The ideal situation would be to have [them] implemented on a chip." Probably not … researchers very rarely do this, and especially not on specific models since chips are expensive to manufacture and difficult to design so they actually yield a performance speedup. <br><br>".. so the runtime efficiency is 100%." Take this out – you can't compute the runtime efficiency of a neuron, and even if you could, we cannot assume that the biological entity is 100% efficient (and at what?!).<br><br>You haven't mentioned yet how you train your supervised networks, or how your situation analysis unit knows how good a certain environment state is. Right now I'm just thinking that magically using genetic algorithms you come up with a quantitative value?!<br><br>In 3.1, you say that you're going to use the tree search depth and number of nodes in a branch which will be searched to the next depth, but didn't you just spend the last section saying how you don't use depth searches like chess? Instead, the neuron just outputs a value? Usually in the tree figure CS guys would write "Leaf node" for a terminal one.<br><br>On page 20, the big-O analysis is flawed … most importantly, why does the tree search depth matter if you're not searching a tree and using neural networks instead (oh, i see – make sure you clearly indicate that this is an analysis of the BRUTE FORCE method rather than the ANN method, something more than a "basic" analysis)? Second, the formula you present is not really very valid – you need to condense the internal equations into a function which provides an upper bound to this.<br><br>In general, it is VERY difficult to perform big-O analysis on neural networks. If you look in the ANN literature, authors do not even attempt this! I highly recommend consulting a computer science professor on the run-time analysis sections, because you definitely need to rethink these. A better approach for this project would probably be directly comparing the run-times of some randomly generated simulations for the brute force approach and the neural network approach, determine a standard way of measuring error based on the "ideal" situation, then rate them both accordingly. You really need to consult a computer science professor on these sections. Be sure to COMPARE the Os after you determine them as well, otherwise why bother?<br><br>I thought it was funny in section 4, that in the introduction you herald your method as being very fast as compared to others, yet you say now that you can't even run your own code with more than one bot at a time! Maybe you should rename section 4, "Expected Results".<br><br>Researchers would use a much more quantitative approach to defining "intelligence" rather than you just looking at the log files. You need a MUCH better approach, because it is VERY likely that intelligence will not just emerge from thin air. Hence, you'll need to tweak your network a lot to get any interesting behavior at all. How will you know what to tweak if it doesn't appear they are acting intelligently at all?! You NEED quantitative measures for defining intelligence here.<br><br>Also, another glaring aspect of your paper is that your results and discussion sections don't contain any results, and hardly any discussion! Didn't you at least collect some data which you could analyze? Do you have anything you can graph? What about run-time data (ie length of time of simulation based on n bots or whatnot)? <br><br>Hopefully these comments help – good luck! And if you need help at all in an entry for next year or even ideas for a project on your own time or help with the math involved or whatnot, I'd be more than happy to help you out with it, or point you toward some good literature to look at in the field so you can be sure to get some good results. Just give me a PM or something.
Sorry if I sound harshly critical below, but I'm comparing it to a paper that a researcher might write in the field (which is what I'm assuming they look for?). If I got something wrong (which is very possibly the case) then please let me know. The paper's largest flaws are the lack of a good introduction and the lack of a good results/discussions section, but read on for more on this. The anonymous poster above gave some great ideas -- heed his words! You state that your research is 'novel', but you never explain why. I am even left questioning why you even looked into this in the first place, and how it may be an important model for the future (use specifics).
In your intro, you mention that your model "retain the basic features that comprise human intelligence" – the field doesn't even know what these are, and frankly the ANNs you're using don't retain them all. (In fact, it would make for an interesting analysis to explain which learning features your model retains and at what gain in run-time compared to for instance Grossberg's Adaptive Resonance Theory model, which does a decent job at incorporating most learning features). <br><br>Saying things like "The field of AI has made limited progress over the past 50 years" is NOT a good idea, since it is absolutely not true! Computers barely existed then! You're basically saying that the entire field was worthless since its inception, so why are you going into it?<br><br>The end of the intro and section 1.1 should be removed completely – they are far too general. Something that REALLY sticks out is the absence of a good introduction. I don't mean to be overly critical, but your reference section is rather short and primarily consists of websites, and your introduction doesn't have a single citation. I bet that they are expecting something along the lines of a literature review in the introduction to illustrate that you've read the relevant literature in the field (academic papers) and that you know how to apply it to make better things. Further, what does your research do that scientists have not done in the past? How is your research unique? Why is it important? Your current intro doesn't even discuss why you chose the networks you are using or even their properties which enable them to learn at all! In general, putting websites in your reference section is not seen at all in the literature due to validity issues. If your references are published in a reputable journal, then you at least know that an expert in the field wrote it. Right now, I get the impression that you just came up with this idea out of thin air for no reason and you're just going to plug it all together haphazardly and hope you get some good results. At the end of the intro you should perhaps say you hypothesis for what you believe will happen (in precise terms) and why, then a short bit on how the results affirmed or disproved your hypothesis (this should also go in your abstract, but make it very brief – most of this will be discussed in the discussion/conclusion section).<br><br>It's starting to sound like a computer game now – WHY did you choose these particular characteristics and actions for the bot? You may want to remove the first paragraph in 2.2 entirely – scientists know all about seeding the weights randomly, it's not novel and YOU certainly didn't model it after observing babies. <br><br>In general, you should always have a caption for all of your figures and graphs as well as a figure/graph caption number (i.e. FIGURE 1. xxxx).<br><br>You're doing a lot of hand-waving in 2.3x with describing the networks in very vague terms. I'm not quite sure you have an understanding of what they actually do or how they actually work, and I'm left wondering why you chose these at all. Your page 13 figure is good (although captionless), but to me it begets the question of why you didn't use an alternate approach such as Markov Processes which predict the future solely based on the past x inputs. Why did you choose neural networks at all? (I'm not saying these would be a better choice, but you never address the issue.)<br><br>"The ideal situation would be to have [them] implemented on a chip." Probably not … researchers very rarely do this, and especially not on specific models since chips are expensive to manufacture and difficult to design so they actually yield a performance speedup. <br><br>".. so the runtime efficiency is 100%." Take this out – you can't compute the runtime efficiency of a neuron, and even if you could, we cannot assume that the biological entity is 100% efficient (and at what?!).<br><br>You haven't mentioned yet how you train your supervised networks, or how your situation analysis unit knows how good a certain environment state is. Right now I'm just thinking that magically using genetic algorithms you come up with a quantitative value?!<br><br>In 3.1, you say that you're going to use the tree search depth and number of nodes in a branch which will be searched to the next depth, but didn't you just spend the last section saying how you don't use depth searches like chess? Instead, the neuron just outputs a value? Usually in the tree figure CS guys would write "Leaf node" for a terminal one.<br><br>On page 20, the big-O analysis is flawed … most importantly, why does the tree search depth matter if you're not searching a tree and using neural networks instead (oh, i see – make sure you clearly indicate that this is an analysis of the BRUTE FORCE method rather than the ANN method, something more than a "basic" analysis)? Second, the formula you present is not really very valid – you need to condense the internal equations into a function which provides an upper bound to this.<br><br>In general, it is VERY difficult to perform big-O analysis on neural networks. If you look in the ANN literature, authors do not even attempt this! I highly recommend consulting a computer science professor on the run-time analysis sections, because you definitely need to rethink these. A better approach for this project would probably be directly comparing the run-times of some randomly generated simulations for the brute force approach and the neural network approach, determine a standard way of measuring error based on the "ideal" situation, then rate them both accordingly. You really need to consult a computer science professor on these sections. Be sure to COMPARE the Os after you determine them as well, otherwise why bother?<br><br>I thought it was funny in section 4, that in the introduction you herald your method as being very fast as compared to others, yet you say now that you can't even run your own code with more than one bot at a time! Maybe you should rename section 4, "Expected Results".<br><br>Researchers would use a much more quantitative approach to defining "intelligence" rather than you just looking at the log files. You need a MUCH better approach, because it is VERY likely that intelligence will not just emerge from thin air. Hence, you'll need to tweak your network a lot to get any interesting behavior at all. How will you know what to tweak if it doesn't appear they are acting intelligently at all?! You NEED quantitative measures for defining intelligence here.<br><br>Also, another glaring aspect of your paper is that your results and discussion sections don't contain any results, and hardly any discussion! Didn't you at least collect some data which you could analyze? Do you have anything you can graph? What about run-time data (ie length of time of simulation based on n bots or whatnot)? <br><br>Hopefully these comments help – good luck! And if you need help at all in an entry for next year or even ideas for a project on your own time or help with the math involved or whatnot, I'd be more than happy to help you out with it, or point you toward some good literature to look at in the field so you can be sure to get some good results. Just give me a PM or something.
Author
Quote:Original post by mnansgar
Overall, it's excellent for high school work!! In fact, I'm very happy for you and surprised that you got all of this hooked up correctly in the first place (congrats!) and that you were able to come up with a plan for analyzing the results! I forsee you as being a great researcher in the field as well, as this is actually a great stepping stone toward some great work. It's clear that you spent a lot of time on it, and that you've also learned a lot from it as well. Many of the students entering that competition actually do their projects for a class or during a stay at a university, so I think this shows GREAT initiative on your part to do it during your free time -- I'm very very very happy for that since very few students will take that initiative if it's not given to them as classwork. If only you were able to understand the networks and the mathematics behind them better than at a very high-up superficial level then you would really have been able to go somewhere with this!
Thanks. I learned a lot from this. I am now attending a boarding school famous for research(we generally win this competition, or place very far). I am taking ResearchChemistry, and I've started the Independent Study class ResearchComputerScience.
I do understand how neural networks work to some extent. My dad did his PhD stuff w/ ANN's back in the very early 90's, so that's how I learned about them. I did a lot of online research about them. I get how a standard multi-layer perceptron works(or at least i think i do), and I understand how error back-propagation works(even a lot of the mathematics, although I'm only taking multi-variable calculus this year, and took single variable last :( ), but I guess I am not so thorough on all the different types of neural networks.
Quote:
Sorry if I sound harshly critical below, but I'm comparing it to a paper that a researcher might write in the field (which is what I'm assuming they look for?). If I got something wrong (which is very possibly the case) then please let me know. The paper's largest flaws are the lack of a good introduction and the lack of a good results/discussions section, but read on for more on this. The anonymous poster above gave some great ideas -- heed his words! You state that your research is 'novel', but you never explain why. I am even left questioning why you even looked into this in the first place, and how it may be an important model for the future (use specifics).
I used the word "novel" because.... everyone else who entered used it, and I felt really stupid because my project wasn't nearly as good(everyone else who entered was a senior). The reason I came to this thread is because I don't know if this is something I should be looking into. I guess I thought it was cool, and built it up in my mind, but I couldn't get ahold of any CS professors(save one, in computer graphics, just for reviewing paper) to help me, so I don't know if this direction of research is even valid. :(
Quote:
In your intro, you mention that your model "retain the basic features that comprise human intelligence" – the field doesn't even know what these are, and frankly the ANNs you're using don't retain them all. (In fact, it would make for an interesting analysis to explain which learning features your model retains and at what gain in run-time compared to for instance Grossberg's Adaptive Resonance Theory model, which does a decent job at incorporating most learning features). <br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Like I said, I don't know as much about the field as I should. <br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>Saying things like "The field of AI has made limited progress over the past 50 years" is NOT a good idea, since it is absolutely not true! Computers barely existed then! You're basically saying that the entire field was worthless since its inception, so why are you going into it?<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Ok, this part was pretty stupid for me. I started getting jumpy as the paper was being written, and didn't have time to thoroughly research the topic. I realize that I should have done a literature review, but remember, I didn't even know what researchers actually did. I just thought you build this awesome thing, and pray it works. Apparently, it works differently ;).<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>The end of the intro and section 1.1 should be removed completely – they are far too general. Something that REALLY sticks out is the absence of a good introduction. I don't mean to be overly critical, but your reference section is rather short and primarily consists of websites, and your introduction doesn't have a single citation. I bet that they are expecting something along the lines of a literature review in the introduction to illustrate that you've read the relevant literature in the field (academic papers) and that you know how to apply it to make better things. Further, what does your research do that scientists have not done in the past? How is your research unique? Why is it important? Your current intro doesn't even discuss why you chose the networks you are using or even their properties which enable them to learn at all! In general, putting websites in your reference section is not seen at all in the literature due to validity issues. If your references are published in a reputable journal, then you at least know that an expert in the field wrote it. Right now, I get the impression that you just came up with this idea out of thin air for no reason and you're just going to plug it all together haphazardly and hope you get some good results. At the end of the intro you should perhaps say you hypothesis for what you believe will happen (in precise terms) and why, then a short bit on how the results affirmed or disproved your hypothesis (this should also go in your abstract, but make it very brief – most of this will be discussed in the discussion/conclusion section).<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Yes the introduction was quite difficult to write for me. The problem was, I had a faint idea of the format of how the paper should be(after looking at professional papers online), but I should have gone to get help in writing it.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>It's starting to sound like a computer game now – WHY did you choose these particular characteristics and actions for the bot? You may want to remove the first paragraph in 2.2 entirely – scientists know all about seeding the weights randomly, it's not novel and YOU certainly didn't model it after observing babies. <br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Like I said, the project was derived from a computer game. Also note the 20 page limit. The copy you are looking at is post-submission editing. The original was very tightly squeezed into 20 pages. <br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>In general, you should always have a caption for all of your figures and graphs as well as a figure/graph caption number (i.e. FIGURE 1. xxxx).<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Yeah… sorry, I didn't format it correctly or something, but there should be captions on everything in origninal submission. I'll check into this.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>You're doing a lot of hand-waving in 2.3x with describing the networks in very vague terms. I'm not quite sure you have an understanding of what they actually do or how they actually work, and I'm left wondering why you chose these at all. Your page 13 figure is good (although captionless), but to me it begets the question of why you didn't use an alternate approach such as Markov Processes which predict the future solely based on the past x inputs. Why did you choose neural networks at all? (I'm not saying these would be a better choice, but you never address the issue.)<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>I guess I didn't want to go into the details of how the networks worked, as I didn't have enough space. The problem was that I didn't have results really, so the entire project falls apart. <br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>"The ideal situation would be to have [them] implemented on a chip." Probably not … researchers very rarely do this, and especially not on specific models since chips are expensive to manufacture and difficult to design so they actually yield a performance speedup. <br><br>".. so the runtime efficiency is 100%." Take this out – you can't compute the runtime efficiency of a neuron, and even if you could, we cannot assume that the biological entity is 100% efficient (and at what?!).<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Never thought about it this way. Good point. <br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>You haven't mentioned yet how you train your supervised networks, or how your situation analysis unit knows how good a certain environment state is. Right now I'm just thinking that magically using genetic algorithms you come up with a quantitative value?!<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Again, I would have gone in more detail, but page limit! I think perhaps the project was very very big, too big for me to handle, and to fully explain this project would take more than 20 pages.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>In 3.1, you say that you're going to use the tree search depth and number of nodes in a branch which will be searched to the next depth, but didn't you just spend the last section saying how you don't use depth searches like chess? Instead, the neuron just outputs a value? Usually in the tree figure CS guys would write "Leaf node" for a terminal one.<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Um… I do search it like a chess tree. Except the Prediction Network is what generates possible moves. Yes, the neural network just outputs a value. However, we train the prediction network against the actual environment state that we record next time.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>On page 20, the big-O analysis is flawed … most importantly, why does the tree search depth matter if you're not searching a tree and using neural networks instead (oh, i see – make sure you clearly indicate that this is an analysis of the BRUTE FORCE method rather than the ANN method, something more than a "basic" analysis)? Second, the formula you present is not really very valid – you need to condense the internal equations into a function which provides an upper bound to this.<br><br>In general, it is VERY difficult to perform big-O analysis on neural networks. If you look in the ANN literature, authors do not even attempt this! I highly recommend consulting a computer science professor on the run-time analysis sections, because you definitely need to rethink these. A better approach for this project would probably be directly comparing the run-times of some randomly generated simulations for the brute force approach and the neural network approach, determine a standard way of measuring error based on the "ideal" situation, then rate them both accordingly. You really need to consult a computer science professor on these sections. Be sure to COMPARE the Os after you determine them as well, otherwise why bother?<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>Yes, I never wanted to do this. However, my mentor(came into project in last 2 weeks), suggested that I do this to try and give some justification for no results yet.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>I thought it was funny in section 4, that in the introduction you herald your method as being very fast as compared to others, yet you say now that you can't even run your own code with more than one bot at a time! Maybe you should rename section 4, "Expected Results".<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>I don't know what part of paper you are referencing… could you clarify? I think this approach would be faster than other "theoretical" approaches.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>Researchers would use a much more quantitative approach to defining "intelligence" rather than you just looking at the log files. You need a MUCH better approach, because it is VERY likely that intelligence will not just emerge from thin air. Hence, you'll need to tweak your network a lot to get any interesting behavior at all. How will you know what to tweak if it doesn't appear they are acting intelligently at all?! You NEED quantitative measures for defining intelligence here.<br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>This was very hard. I couldn't think of any good ideas. Could you suggest some perhaps? Even my mentor couldn't come up with any good ideas.<br><br>Tweaking the network would be near impossible(we're talking 100x100 ANN minimum). Tweaking weights in that is suicide.<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>Also, another glaring aspect of your paper is that your results and discussion sections don't contain any results, and hardly any discussion! Didn't you at least collect some data which you could analyze? Do you have anything you can graph? What about run-time data (ie length of time of simulation based on n bots or whatnot)? <br><!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br><br>TAH-DAH!!!! You just discovered why my project was complete garbage! I was pretty depressed after submitting my project, as I didn't really have results. Quoting my dad, "You don't publish to publish, you publish results!" :[<br><br><!–QUOTE–><BLOCKQUOTE><span class="smallfont">Quote:</span><table border=0 cellpadding=4 cellspacing=0 width="95%"><tr><td class=quote><!–/QUOTE–><!–STARTQUOTE–><br>Hopefully these comments help – good luck! And if you need help at all in an entry for next year or even ideas for a project on your own time or help with the math involved or whatnot, I'd be more than happy to help you out with it, or point you toward some good literature to look at in the field so you can be sure to get some good results. Just give me a PM or something.<!–QUOTE–></td></tr></table></BLOCKQUOTE><!–/QUOTE–><!–ENDQUOTE–><br>[/quote]<br><br>They helped a lot. Sorry if I sound a bit defensive, but I have so much homework that I'm going to explode. I was hoping that perhaps you could serve as a pseudo-mentor? I am entering this in another competition in December(PAPER could use help!!!! :) ), and going to try for ISEF(International Science and Engineering Fair). <br><br>My other questions was: It seems that mainly biology and chemistry projects win. My dad, who works in CS field, always pushes me towards bio/chem projects, as they are the ones which win. I am doing Research Chemistry this year(I am interested in this, but not as much as CS, and my project will be using computer to model some phenomina). What do you have to say about this?<br>
Maybe this is a stupid question, but why didn't your dad look over this if he actually did research in the field?
>> It seems that mainly biology and chemistry projects win.
I've heard this as well, and I've been told and believe that this is because at the high school level, even advanced students are typically not mathematically mature enough to understand modern research, muchless come up with novel research in the field. Also, it's easier to come up with good experiment ideas in for instance biology since so much is unknown. In math and computer science, you really have to know the field first before you propose a project. If you feel that you can understand the math and abreast of developments in a math field, then I have no doubt you'll have similar opportunities to win.
>> I was hoping that perhaps you could serve as a pseudo-mentor? I am entering
>> this in another competition in December(PAPER could use help!!!! :)
I'd be glad to assist you if you need some help. I'm just a student myself, going into Computation and Neural Systems for graduate studies (making mathematical models of the nervous system derived from psychological and neurobiological data). I currently take about half of my courses in physical brain science (biology (cell/neuro), psychology (cognitive/psychobio), and mathematics (cognitive neuroscience/logic)) and the other half in simulation (EE (digital electronics/signals and systems)), CS (AI)) for some very hectic semesters, but I have an excellent background in research and graduate coursework so could possibly be of some help.
For the paper though, you can improve the introduction and methods portions a lot, but unless you're willing to run more experiments, you can't do much with the results. Many ANN papers are just descriptive due to the nature of not having exact solutions for the involved differenetial equations, so you may be able to pass it off as that IF you have a good rationale for choosing your networks and some results. You also need to condense your writing a lot and fit everything in -- right now, the majority of it is fluff, so pack in the technical details! I know 20 pages seems like too few, but you'll have to condense everything into in some cases 3 page papers in the future when you publish, so get used to it! The general idea is that if someone can't replicate your experiment exactly from your description, then it isn't a good description. I should have no question in my mind about how you implemented this and why when I'm finished.
>> Tweaking the network would be near impossible(we're talking 100x100 ANN
>> minimum). Tweaking weights in that is suicide.
It's not the weights you'd tweak, but things like the learning alogrithm and the threshold values (or functions). Like I said, it's HIGHLY unlikely you'll get any learning at all by just smashing all of this together, and you'll need guides to figure out what went wrong. If you can't figure out what's wrong, then again you don't have results. I can't think of a quantitative error term off the top of my head since your system is rather arbitrary, but you'll have to find one or else you won't have a guide for what went wrong and why things aren't learning.
>> I used the word "novel" because.... everyone else who entered used it
Well, I agree that it's a good term to use, but only if you can explain why.
>> It seems that mainly biology and chemistry projects win.
I've heard this as well, and I've been told and believe that this is because at the high school level, even advanced students are typically not mathematically mature enough to understand modern research, muchless come up with novel research in the field. Also, it's easier to come up with good experiment ideas in for instance biology since so much is unknown. In math and computer science, you really have to know the field first before you propose a project. If you feel that you can understand the math and abreast of developments in a math field, then I have no doubt you'll have similar opportunities to win.
>> I was hoping that perhaps you could serve as a pseudo-mentor? I am entering
>> this in another competition in December(PAPER could use help!!!! :)
I'd be glad to assist you if you need some help. I'm just a student myself, going into Computation and Neural Systems for graduate studies (making mathematical models of the nervous system derived from psychological and neurobiological data). I currently take about half of my courses in physical brain science (biology (cell/neuro), psychology (cognitive/psychobio), and mathematics (cognitive neuroscience/logic)) and the other half in simulation (EE (digital electronics/signals and systems)), CS (AI)) for some very hectic semesters, but I have an excellent background in research and graduate coursework so could possibly be of some help.
For the paper though, you can improve the introduction and methods portions a lot, but unless you're willing to run more experiments, you can't do much with the results. Many ANN papers are just descriptive due to the nature of not having exact solutions for the involved differenetial equations, so you may be able to pass it off as that IF you have a good rationale for choosing your networks and some results. You also need to condense your writing a lot and fit everything in -- right now, the majority of it is fluff, so pack in the technical details! I know 20 pages seems like too few, but you'll have to condense everything into in some cases 3 page papers in the future when you publish, so get used to it! The general idea is that if someone can't replicate your experiment exactly from your description, then it isn't a good description. I should have no question in my mind about how you implemented this and why when I'm finished.
>> Tweaking the network would be near impossible(we're talking 100x100 ANN
>> minimum). Tweaking weights in that is suicide.
It's not the weights you'd tweak, but things like the learning alogrithm and the threshold values (or functions). Like I said, it's HIGHLY unlikely you'll get any learning at all by just smashing all of this together, and you'll need guides to figure out what went wrong. If you can't figure out what's wrong, then again you don't have results. I can't think of a quantitative error term off the top of my head since your system is rather arbitrary, but you'll have to find one or else you won't have a guide for what went wrong and why things aren't learning.
>> I used the word "novel" because.... everyone else who entered used it
Well, I agree that it's a good term to use, but only if you can explain why.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
