Readings
Diana Applegate
I found the discussion of evolution as search to be really interesting. Since GA's and GP are analogous to some of what we learn in bio class, I was able to easily comprehend these processes. You mentioned in class that there are lots and lots of applications in this area, and I'd be quite interested in hearing about some of them. It was sort of surprising to me at first that the multiplication and manipulation of our chromosomes (what makes each of us who we are!) could be so similarly modeled in programming. More evidence to assert that the body is simply a machine?? Who knows...
Our lab for the week, which involved watching movies, was great fun as well. The videos were informative and also very entertaining. I especially enjoyed the tour of the MIT Media Lab. Anyone up for a road trip?? ( : I'd love to check out that lab someday. I guess I'm one of those people that the lab is trying to target for financial support...I really enjoy checking out demos!
I don't have much else to say about this week...I'm starting to panic a bit about the programming project since I feel that my CS110-only background is insufficient. Does anyone else out there feel like they don't know where to begin? But it's something that I'd like to attempt, at any rate. I'll definitely be stopping by soon to discuss the details!
Jocelyn Arcari
I find it so interesting that we carry over terms and ideas from other fields to be used in a seemingly unrelated subject all-together, however, the use of "chromosomes and populations and mutations" all make wonderful sense as terms. Our discussion in class about the question of our "intelligence" and whether or not we have created the whole idea was unsettling in a way. We don't like hearing that perhaps we are not as intelligent as we think we are. However, trying to put biases aside, I find it hard to believe that our "intelligence" is no more genuine than that which we can program. I still think that feelings create a huge barrier between us and computers and that feelings are very real. Maybe the whole difficulty comes back to our interpretation of intelligence. If we question whether or not computers have any sort of intelligence of if they could, I think that they do. But, it is definitely not the same type and speaking about it as if it were the same is what becomes unsettling. While many of our actions and functions are program-like, not all of them are.
I wish Nilsson had gone into a little more background on genetic algorithms. He did speak of genetic programming and did a good job at describing the program for the evolution of a wall-following robot. This reminds me -- is anyone participating in that robot contest? I would like to get more information on the paper we read about BDI ARchitecture for Embedded Rational Agents. Do we get to see any of the results of your work? Any robot models? It was interesting how carefully words had to be defined such as "reasoning," "mental acts," and "control acts." Finally, can you go over the Semantic commitments, architectural commitments, and epistemological commitments for the SNePS BDI Architecture a little more? I'm not sure I'm getting everything out of that section that I should.
David Costello
I found last week's lectures on genetic programing very interesting. I was wondering if you could give further examples of it's applications in music. Has a good song ever been created using this algorithm or were you just theorizing the future uses of genetic programing? However, I was less impressed by Moravec's chapter 4 analysis which I thought was more about science fiction than true science. It seemed to me a little premature to theorize about artificially intelligent robots in space when we can't even create artificially intelligent robots which fly in the air (last Monday's movie). Is copying a mind into a computer researched in AI or is it too advanced to even speculate about?
Sonia Dubielzig
Watching the movies for robot lab really allowed me to see how artificial intelligence has progressed from, say 10 years ago, when Moravec wrote Mind Children. Was the idea of wearable computers acting as an extension of one's brian, like the cyborgs wear and like Moravec discusses in the Symbiosis chapter, being kicked around in Moravec's time, or did he suggest it? Certainly, the cyborgs and the shoe-computers seem to bring his chapter to the present. It also raises the question of consciousness and idea of self again; is a cyborg's computer really an extension of the brain? I can't believe that professors allow them to take their tests with their computers on; the computers should be considered an aid, not a part of one's consciousness. Or should they?
I certainly think that it will be nice to have computers that can use artificial intelligence to better relate to human beings. The inconvenience of having to GO somewhere to check one's e-mail or get on the internet would be eliminated. On the other hand, it would be that much easier to procrastinate.
The genetic programming that we discussed in class on Thursday also highlights one of Moravec's assertions, that computers will be able to "evolve" and "adapt" into amazingly complex beings, unleashed from the slow plodding pace of biological evolution. Having programs figure out for themselves what the best program is allows the computer to go beyond what the human mind can think of, and opens up the possibilities for creations not in our comprehension now. Perhaps computers could use genetic algorithms to configure a better model for themselves, or robots could "think up" a new type of robot.
Benjamin Flynn
I've known what I'd like to write this response on, so I'm a little disappointed that I've put it off until now, but I suppose the extra time composing my thoughts may not have hurt.
It occurred to me that one person who forsaw the way artificial intelligence would be incorporated into the home was Ray Bradbury. When in doubt, look to Ray. HIs short story "There Will Come Soft Rains," from the Martian Chronicles, tells of a house that continues to perform its function, even after its owners are gone. Each unit of the house functions independently, there is no "robot butler;" individual appliances perform their function as part of the whole. I think it would be worthwhile for all of us to give it a read through. :)
I have also encountered quite a bit of strong reaction to the Cyborg's of MIT. Many gave the Poor Tom "I'm going to run off and live in the wilds" reaction. The idea that people are going to have text readouts displayed before their eyes is widely upsetting. I myself feel a bit dismayed, although I realize that a proper cyborg system could be incredibly useful, and faciliate understanding and propper functioning in highly complex situations. One of my own biggest worries about this system, however, is that we would become like the Borg of Star Trek, relying ever more heavily on a central database to shape our perception and opinion of the world. Even if the database only contained "factual" information, we are all aware of the difficulty in creating true objectivism. Even to the critic, television has a strong influence, arguably perhaps but I think there is sufficient grounds to say as much, over the way we perceive reality. We associate what we see on TV with real life. To an extent, this is necessary for entertainment, a "suspension of disbelief." However, it is difficult for us to decisively separate the "real" from that which we glean from the TV. I have a whole theory on this, but I won't bore you or the class with it. :) Anyway, my point is that the cyborg unit could act as a radically more pervasive TV, shaping our view of the world, not necessarily in a beneficial manner.
Emily Greenfest
I have been much interested these past few weeks in our discussions of searches. Having examined many AI related problems in an earlier Machine Learning class I was not surprised when the point was raised in class that many AI problems boil down to searches -- in fact, it seems quite logical that they do so. Even at the level of our non-planning gridworld robot, if that robot has a goal a search is being performed. The robot is searching the space (albiet not necessarily in any directed or systematic way) for the goal. The same is true of our last few labs and lego-robots: we are asking the robots to head for a light source and to do so they must find (search) an appropriate pathway to obtain their goal.
I find the idea of searches very interesting, especially when they are applied to Genetic Algorithms. A GA is, at it simplest, a search method. It allows for a non-random search to be performed (i.e. successive stages in the search do relate back to previous stages) while at the same time minimizes the chance of the search being stuck in a small portion of the search space by adding a touch of randomization to the process. I suppose that makes little sense, but if you play with them enough, it should become clearer. the point is, that it is a search. And, more importantly, it is a search based upon genetic principles -- the same principles that the biologists claim are in charge of how organisms evolve and change. It is because of this relationship that the workings of GAs raise some interesting questions: it has been long been claimed (well, for the past 100 and 40 years) that evolution is not directed: it has no goal. And yet, with the implementation of GAs (same principles, remember) we have seen that for genetic recombination to work effectively it must be directed in some way (via the fitness function). And for that direction to be effective one must have some notion of a goal, however vague it might be. One must have some notion of how one individual might be better than another. It is arguments like these -- the anti-directed no supreme being goal arguments -- that have led to many of the changes in evolutionary thinking these past 10 - 15 years. The idea that a genetic process might be a semi-directed search has suggested that perhaps, there is no selection in evolution, or there is no fitness evaluation.
These are interesting questions, and I think they argue, better than most, how AI is really one of the most multi-disciplinary sciences: not only have been GAs (and other searches) been used to solve problems in physics, math, social sciences, medicine, etc., but they also pose questions and problems for other sciences including psychology, biology, etc.
Im a little intimidated by this program maybe you can talk about it alittle more in class. First of all, where do i start?
I recall that in class you mentioned that we should just start off simply ignoring the search procedure that we intend to use for our problem solving.
I know I shoudl create a board class. This class must somehow incorporate the fact that the board has an 8X8 interface, there are two types of pieces (black and white), and the pieces should be designed in such a way that they can be indiviually accessed. Am I on the right tract here? I think I can manage creating the board but i am confused at the prospect of a possible moves table. What is the algorithm that generates the possible moves? Do I have to use one of the recent concepts learned in class to generate such an algorithm?
Also I was wondering is there going to be robot assignments along with this project? The reason that I am just now starting to generate the pseudo code for this project is because my group and I have spent many hours last week debugging and perfecting our robot for today's lab.
Sarah Klaum
Monday night I experienced once again the excitement that precluded my introduction to AI. Observing the technology that was employed in the tennis-ball snatching robots, and those that created maps and searched out appropriate rooms was impressive. I found myself cheering for the competitors in much the same way I would have for humans, a clear indication to me that I indeed saw in them some kind of "intelligence." I'm curious as to what advances have been made with some of the other robots, for instance the one that was being trained to take a hockey puck from a human's grasp. (I've forgotten the name of this promising specimen). After observing both this robot and the robot that learned to take its first steps, I wondered when and how aspects of memory would be employed, so that they wouldn't have to complete the same tasks each time they were turned on. (This may in fact have already been an integral part of the program not shown in the demonstration, I just don't recall any mention of it).
I was amazed and a bit alarmed by the prospect of a "body net," as one of the presenters labeled it--an insert for the shoe that could turn one into a walking and communicating data base. (We are already, of course, but the idea of shaking hands with someone and having personal information appear on a computer screen is not something I had explicitly thought of before). The "wearables" also brought back to mind a question that I have struggled with before--where do we begin to "lose touch" with our physical reality and environment as we now know it? I was troubled by the student who wore the large helmet/head gear and other equipment--it appeared that at times he relied on his computer to do facial scans so that he knew who was approaching. I am excited by the thought of technology that can enhance our understanding and appreciation for our surroundings, but I disliked seeing such a reliance on a computer for an observation that can be made with our own natural capabilities.
Frank Rusch
Genetic Algorithms are of particular interest because they model how living things develop and evolve. In the example of the wall-following robot in grid world, the spaces next to the wall are like food sources, where the most fit individuals are the ones that procure the most food. Because the evolved program is looped many times, a truly fit program will develop a sense of what to do in every possible situation. I wonder if Genetic Programming is as effective as neural nets, since with neural nets, the weights are directly adjusted according to whether they produced the right output. With GAs, the alterations are random. In other words, two parents who have programs with some good features will not necessarily pass down the good traits any more than the bad traits. So the process takes longer, and requires more steps. Because these evolutions are processed in computers, we can get them done in a reasonable amount of time. One great advantage over neural nets appears to be that the progamming code itself is directly modified, so there is more freedom in the implementation. This code still has to be compiled, so it's not as if the computer's "genetic materia"l (1's and 0's) is being directly modified.
When looking at the "most fit individual" on page 68 of Nilsson, I notice some apparently redundant statements (e.g (NOT (NOT (NOT etc....))))). This makes me wonder if humans and other creatures have evolved in such a way, where simpler means would reach the same end. Those extra evaluations of functions might not make a difference for computers, who process the instructions very rapidly. For living things, time is more of an issue.
First of all, I must say that I greatly enjoyed watching the videos in lab last Monday. This was an excellent opportunity to find out about some relatively recent advances in robotics and other AI areas. Also, it was fascinating to see the ideas and concepts we learn about put into practice. Learning, planning, subsumption architecture were only a few of the concepts from class that came up in these videos. As I was watching the clips from the robot competition, it struck me how much emotional involvement there was from the spectators (including our class). It is understandable that the competitors are anxious to have their months or years work rewarded with a good robot performance, but at moments it seemed as if they (and some of us) were cheering for their kids in a junior league game. I guess we are very hesitant when it comes to identifying intelligent behavior in these machines, yet we do not seem to have the least problem with personifying them.
In the robot competition program, I liked the idea of using two robots to perform the required task faster. This reduces the amount of work that a single robot has to do, but, as the robot designer claims, introduces an additional complication, namely communication between the two robots. A similar idea was used in Sam, the MIT robot, (I hope I remembered his name right), whose body parts were capable of acting and learning on their own, as well as in collaboration with the rest of the body. I guess a system consisting of several semi-autonomous agents significantly cuts down the reaction and planning time for the agent, as long as the agents are not overly dependent on each other.
I find the genetic programming and algorithms very attractive topics, which probably deserve some more attention. I would be very interested in learning more about them, or at least looking into some of their applications in AI. There is a neat question at the end of the chapter, about the meaning of the biological terms genotype and phenotype as applied to GP. Genotype are genetic constituents of an individual and phenotype are the visible properties of the individual, arising from the interaction of genotype and the environment. So, the genotype in GP would consist of the program code, and the phenotype would be the performance of the program given a particular set of circumstances. For example, in the case of the wall following robot, the same or very similar codes (genotypes) might result in very different behaviors (phenotypes) if we change the position of the robot in the environment. It seems that all of the GAs and GPs are used to perform tasks which do not require speed or excessive precision. Have I understood this correctly, or are there problems where these methods produce a fast and optimal solutions?
I've a spent a bit of time thinking about the question of intelligence, and when an object or being should be considered intelligent. The idea has come up that advances in AI might serve as well to demystify the notion of intelligence in humans as it does to advance the intelligent behavior of machines. And in turn, that the demystification of intelligence will somehow negate it or disprove its existence. I do not believe that this will be the case. I don't think that intelligence needs a mysterious or magic component in order to be considered real. Say, for example, that we were to discover that feelings such as happiness were just a bunch of certain surging electrochemical. Would this prevent people from becoming happy? Or say that we could exactly map the brain's function when one smells one's favorite food cooking. Would that stop that person's enjoyment? I think this would make intelligence all the more miraculous, that the human mind could produce so much given so little. Maybe we should consider intelligent behavior in more quantifiable terms and be content with a simpler recipe for human intelligence.
Now, I've almost just about had it with the Moravec book. With each successive chapter it seems to get more and more into the scary, incomprehensible-to-us-mere-humans doomsday scenario. Yeah, yeah, I get the idea, that machines are more efficient and faster, and might as well take over where we can't. And we'll be too slow and stupid to grasp all the amazing things they're discovering in outer space and all that. Is this really a world that we're even equipped to care about? I can't decide whether being a mere dumb human in this future is going to be really really depressing or if it will bring on a whole lot of apathy. I think most of us even today prefer to associate with our own circles of friends, who share our motivations and understandings. Why would we *want* to relate to super-intelligent robots? Let them do their own thing. The patch of moss doesn't give a darn about what we humans do or discover, and it still seems to have a fine old time, prospering in its mossy way. The apes that we evolved from don't ever come back to us and ask us to explain gravity and relativity and so forth.
Emily Sweeney-Samuelson
I loved watching the videos on Monday. It was very informative to see robot competitions and up-to-date applications of AI. I only wish there had been more history given of each group that entered the contests; more detailed descriptions of how their robots operated and narratives of the process they went through in building their robots and what problems they ran into. Of course, then the videos wouldn't be as enjoyable for a broad audience, but I wanted to hear more technical information, so I could better understand the behaviors I was seeing on the videos.
The intelligent van was fun to think about. I have seen cars with those navigational systems, and I know how disconcerting it can be when your car talks to you. I wonder how soon these cars will be common possessions.
I'm glad we didn't completely skip the chapter on machine evolution, because it seems to be a very important topic with lots of possibilities in AI. It does seem to fit in the course better now that we know more than we did after chapter 3, though; I read it back then and it didn't sink in as well as it did this time.
For some reason it is very strange and a little eerie to read about mathematical/computing models of evolution. It seems normal to me when I read about robots imitating biological processes like vision, and even neural nets don't make me cringe, but something like evolution being duplicated by machines is almost scary to read about at first. Maybe this is because it is a more abstract or broad phenomenon, not a specific function of one being, like vision or brain functions. It makes me think of the race of robots that Moravec talks about, which is certainly chilling. When the more technical and specific aspects and techniques are discussed, this feeling wears off somewhat, because it becomes more defined, not unknown, but the general concept is still a little hard to swallow, even when I think about its wonderful applicability to AI and exciting prospects.
Perhaps it's even harder to swallow when I think of some of those very advanced prospects, because the thought of a machine procedure that compares very closely to the evolution of a natural species is a jarring juxtaposition.
Tim waring
I have read Kevin Kelley's book out of control This book is called the "rise of the neobiological civilization." The idea behind this 500 page book is that machines and biology are coming to a head. Machines are becoming more and more life like and life is becoming more and more engineered. He spends a great deal of time talking about the idea of machine evolution, and about a lot about other AI strategies, including sumbsumption architecture. This chapter in Nilson (whom kelly interviewed in Out of Control) is informative. Kelley talked about some drug companies that are begining to use machince evolution, specifically genetic algorithms to find solutions to otherwise unanswerable questions of the drug world, like what molecule inhibits cytokinesthetase, or promotes the production of pyropathsilirone?
I have been taken with this idea too. Since freshman year of college I searched the web for shareware programs that enable my mac to be a small ecosystem, with little autonomous agents living and evolving within. MacTierra and a couple of other good ones are out there, but becuase i didn't actually write these programs it's hard for me to completely enjoy them, as i don't have a full understanding of what they are doing.
Sorry to switch back to submsumption architecture but i like it and i am thinking about it more as we work on modifications to our robot. It strikes me how easy it is to add features to a s. a. agent. simply build a paralles process and include parameters to handle it in the arbitrator. Kevin Kelley talked to great hights and lengths about explaining the easy nature of adding behavior, and now i see it, it's great!
Sarah Waziruddin
I really liked the robot movies we saw last Monday. It was exiting to see recent developments in artificial intelligence. I also liked seeing the development, or evolution, of the self-driving car. The technology developed there is very useful.
Nilsson chapter on genetic algorithms was interesting also. In this chapter, Nilsson sounds more and more like Moravec, likening machines to humans by mentioning evolution and selective survival. Functional programming sounds really cool and it would be nice if we could have a chance to look at or to actually program a small lisp program. The tournament selection process sounds familiar to tournaments in ancient times-- each person fighting for survival. These randomly generated programs are also fighting for survival. Finally, a wall following robot is created as a result of evolution.
Moravec wrote about the use of artificial intelligence in programming-- he hinted at the idea of machines being so evolved that they would understand our commands in English and there would be no need for programming languages. The process of the evolution of these programs is similar to that idea. The final wall following program wasn't actually programmed by us, but was a result of intelligent behavior.
Leslie Zavisca
The most interesting part of my week, with regard to AI, was seeing ANTZ with my dad and being able to explain quite a few of the technical aspects that he had questions about. It was a fun movie. Go see it if you get the chance and if you like Woody Allen.
Speaking of movies, last week's lab was very intriguing(especially since we got the chance to see Alan Alda walking around with a computer on his head). I had no idea the kind of research that is going on at MIT. I would love to go there and see it myself. For me, the most memorable display was the acrobatic robot. I hadn't realized just how advanced AI is becoming these days. It is quite amazing. Another story involving my dad...I was in the King of Prussia mall with him this weekend and in one of the hallways was a new Acura with the satellite linked navigation system. I am curious as to how sophisticated this system really is. I can't help but be skeptical and think that it's got a very limited knowledge base and can only assist with very basic tasks such as getting from city to city but not from block to block. Does anyone know more of the specifics involved with this system?
I was also very interested in our class discussion, prompted by Emily, on intelligence and the possibility that we are simply incredibly self-absorbed robots overestimating ourselves. Pondering such an idea can be pretty humbling, but I find myself asking that same question all the time, especially after studying behavior analysis last year (another tie-in to psychology). Both fields, as far as I've studied, are very stimulus-response oriented, leaving little room for emotions or true choice.
As for our robot, meta sensing proved to be quite a learning experience. Our first draft of syntactically correct code looked perfect to us and should have worked just fine, but because we're using the code on an actual physical being, we ended up running into problems related to the length of time certain motor commands take to complete or the fact that the bump sensors are often depressed for more than just a second... These conflicts reminds me that theory and reality often disagree--all code is good code until proven otherwise by unexpected obstacles. Also, we had already randomized our robot's BACKthenLEFT and BACKthenRIGHT so we're having trouble actually getting her stuck in a corner. We know that our JAM function works because we've tested it by manually bumping the sensors, but we haven't really gotten her stuck in a real corner yet. We'll proceed with testing her tomorrow...