Readings
I really enjoyed the Konane tournament last week. Although I am not a good strategic game player by any means, it was great to be able to interact with different programs. I especially enjoyed Peter's program, even though its comments were a bit out of line at times! ( : But that made playing all the more entertaining. Congrats, Peter! Next, I appreciated the extra review of examples of resolution in FOPC. Each time we do an example, the process becomes more second-nature to me. As I've stated in just about all of my recent responses, I'm looking forward to delving into Natural Language. I had lots of fun experimenting with Eliza, and the "arm and a leg" vs. "nominal egg" example in class was quite amusing. I've been accused by many of having a slight NY accent even though I live in New Jersey. It seems to shine through when I say "water" and "orange". Accents are of continuing interest to me, as are creoles.
Jocelyn Arcari
Well, class on Tuesday showed me that I definitely need to practice converting wffs to clauses and applying resolution. It seems like we did so many steps that it would be hard for me to recreate them on my own. I will have to re-practice the example we did in class.
The course web page suggests that we may wish to discuss our over-all experience in this class, so I thought I might address this. When I came into this class, I came with definite thoughts that computers are not able to think. Throughout the course of the semester, I inched farther away from my stance and started to consider the possibility that I was wrong. Some barriers that remained for me (for instance, in lab) was the trouble that if we were actually programming the robot to do something, how could it possibly be thinking on its own? When we saw things such as emerging behavior that was not specifically programmed into the robot, I found that interesting but it was still fairly easily explained away by other aspects of the program. Having the opportunity to watch more sophisticated robots on the different videos you showed us helped bring to light the possibilities of advancements in AI and the possibilities of robots becoming more intelligent.
I'd be curious to know how many people came to change their opinion about whether or not computers are able to think after taking this class -- either moving farther away from believing that they CAN think, or moving closer towards this belief??
I also am still hung up on the question of whether computers need to process information in the same way humans do in order to be called intelligent. I come back to my project on Deep Blue and the conclusion I made that Deep Blue is not intelligent at all. It simply houses tons and tons of information and can sift through its database to figure out what move might be a good one compared to all the moves in past chess games. But, to an outsider unfamiliar with the processes going on, it might look intelligent since it was able to beat the world champion in chess. Still,I cannot call this machine intelligent, since it does not "learn" from its opponent or from its mistakes mid-game. If it had this capability AND were able to use its large database, I would reconsider my answer. I have come to the conclusion taht I do not think that a machine needs to mimic human thought processes (that we do not even fully understand) in order for us to call it "intelligent," but I do think it needs to have the capacity to learn. Character recognition programs like those used for our postal service do seem (almost) intelligent to me since it can sort things even though no two people have the same handwriting. That is such an amazing advancement that achieves exactly what I think computers should be used for -- doing things that make life for humans easier. Calculating huge numbers, storing large amounts of information, searching through information and answering a query, and any other result that makes life more manageable is what I think the real benefits of AI research are. Even if we decide computers are just as intelligent as humans after this class, I do not fear the day when computers will take over the world. I do not forsee them being able to run our lives, rather, being able to Help us run our lives more efficiently.
This has been a really enjoyable class. As you probably can tell from my disorganized thoughts that I still have many questions to figure out for myself, but all of our discussions have been very interesting this semester and I'm glad we had the opportunity.
David Costello
I have some questions concerning the final project. Do we need to submit a report after the presentation on Monday or is the presentation the only thing graded? If there is a paper to be submitted what should it include? Is the paper due on Tuesday or is it due on Monday? Also I was wondering if there is a time limit for each presentation.
Another set of questions I have concerns the Konane tournament. What methods of static evaluations won in the tournament? Did deep depth searches win more than shallow depth searches? Also, if a Konane game has shallow depth searches yet a winning record does that imply that it's static evaluation is vastly superior to itís opponents or can it be due to another factor?
Sonia Dubielzig
Reading through the _Natural Language Understanding_ packet, I began to think about the complexities of our language that we don't notice. Psycholinguists like Noam Chomsky propose that all humans have language ability "hardwired" into their brains automatically--children are actually _looking_ for a language from the time they are born, when they begin to listen to noises and make them themselves; through the time they begin speaking, using untaught language rules like saying "mouses" instead of "mice".
I remember reading an article on language formation which recounted how a school of deaf children in central america had created its own sign language. At the school's inception, each of the students had developed their own signs to communicate with their own families, and the sign language language they developed to communicate with each other was very primitive. However, after two years, the language had developed into a complex sign language of itself; as each new wave of students came in, they filled in the gaps, creating verb tenses, agreements, adjectives, etc. The older students, who had known the language longer, were not as adept at using it as the younger ones who learned it quickly. Basically, this deaf school had its own language, created by a group of children who had never heard any other language.
If language ability is hardwired into us so that we cannot consciously understand how we learn it, it seems that it would be very difficult to program straight, abstract rules of grammar and syntax into a computer. The neural network model of language learning that we discussed in class on Tuesday seems the best solution to solve a problem we do not fully understand ourselves. Why not let the computer do as children do, and look for the rules by itself?
Hello. Over break I got a chance to look at Konane once again as I showed off my program to a few friends. One of them is particularly good at that type of game, and beat my ai player at search depth 5 on his first try. After he had achieved victory, we spent a while discussing what he was doing when he was decision making. We tried a few simple strategies to change to evaluation function, but none proved as successful as the original. We thought that what was really necessary was not to look only at the number of moves available at a given turn, but to look for certain patterns in the way pieces were lined up. Unfortunately, the slow connection from his house made implementing this sort of suggestion impossible at the time.
I have more recently been turning my attention to my final project. I have an on paper layout of how it should work -- I need to look a little closer at back propogation to make sure I really understand how and why it works. I'll probably do some of that later tonight.
Peter Ingebretson
I was pretty impressed with the lisp program we saw a while ago, but after seeing how resolution works, I'm a little surprised that you, and the other designers of that software, decided not to use a more "natural" reasoning method instead of resolution. Although it is attractive to see a program that reasons in a manner similar to the way that we reason, from what we've seen, resolution can be interpreted as generalization, or simplification, of the way that we reason.
I am also very interested in the natural language understanding packet that we read, and I hope that we will spend some more time talking about language understanding. I know that some expert systems can understand one topic to the degree that they behave as though they understand the topic and questions that a person might ask about it. One approach that natural language systems seem to take is down this route, towards an expert-expert system, one that can understand different topics, but is limited either to a finite number of subjects or to a finite number of modes of discourse about different subjects. To me this doesn't seem like a very useful approach in the long run, although I'm not sure what a more feasible alternative might be.
Sarah Klaum
It is easy to see why natural language understanding, with all of its related diciplines, is an AI complete problem. Whenever reading outlines of syntax and grammar, I am continually amazed by a child's ability to assimilate and comprehend language. Many of us have had lessons in english grammar in grade school or junior high, yet ever since a young age knowledge of syntax has been intrinsic. Linguistics and psycholinguistics are both interesting diciplines, in my opinion. I am curious about what AI/computational work has been done to model any of the theories proposed by either of these diciplines about how we _learn_ a native language, not just how we understand it, given a grammar. I would imagine it would be a difficult task, given how much we still have to learn about the acquisition of a first language, not to mention how this influences learning other languages.
Wow, so things are winding down now, so I've been thinking about where exactly we've gone and all those "what the future holds" questions that real AI scientists don't think about because you can't make decent predictions about the future and you shouldn't anyway since they probably won't come true and even if they do, what does it matter? Phew.
I was talking to Steve Lindell about the cognitive sciences (mostly psychology, but AI connections crept in there at the end) and he made the point that a big division between psychology and other sciences is that in psychology, even though the number of possible subjects to study are enormous, each one is only so "deep". That means you read up a bit on it and it's not long before you run into the big "I don't know". There's so much about the human psyche (and any sort of cognition really, except for the very basic, e.g. insects and whatnot) that we don't understand. The big "I don't know" means "we observed this and this and this and this is what we think about that, but as for anything else, we're not really sure. We're still experimenting and maybe in 10 years we could explain why humans do this, or how group consciousness fits into it all, or why we're so this and that and..."
One of the reasons I've liked computer science is that there's a certain compactness to it all. I mean, there's several different fields of research, but they aren't vastly different. They usually involve a lot of programming and/or math, or in the case of engineering: wires and silicon. All except for AI. AI is sort of the odd man out in the group. Most other areas of computer science are narrow, but deep in the sense that they are well understood. Of course, that makes sense since WE'RE the ones who developed CS, so we ought to understand our own creation.
But AI--AI is almost the engineering side to cognitive science. The fields are many, but they're shallow in the sense that they're not well understood. There's the big "I don't know" hanging over what's still being done. Bit by bit that "I don't know" is being eroded, but it's slow. Even if the goal in AI isn't to necessarily "do" the ways humans "do", the thing I've always known about a good algorithm is that you have to have a good understanding of a problem and how to accomplish its solution. A theory is only so good as the program it produces. So maybe we don't have to figure out how humans accomplish a task--we still have to know how that task can be accomplished and that's the hard part.
But in a way, that's what makes cognitive science and particularly AI so very interesting--the solutions are so rewarding, especially for being so hard-won. Who doesn't like a good mystery after all. And the most extraordinary and unique ideas will emerge from AI in the future, I feel certain, to say nothing about what we have now.
Hmm, maybe I'll reconsider graduate school... :)
Frank Rusch
Now that the Konane tournment is over, I was thinking about the difference in the programs. The static evaluation function was the main thing we had a choice about. What turned out to be the best static evaluation function?
About natural language understanding-- It looks to me like a language-understanding computer could not be static; it would have to have a memory to recognize context, since a single sentence could have different meanings depending on what was preceding (or inflections, in spoken language). I think context could also depend on the person talking, and the typical style of their communication. Also, people can understand language, even if it doesn't follow the rules of grammar.
When a computer decides that it wants to talk about something, how does it decide? The Eliza program will not say anything concrete about whatever you ask it about. Eliza comes with pre-structured sentences built-in, with the keywords substituted. Learning to write a language involves studying the grammar, and listening to how the language is typically used. when a person is about to say something, they don't generate random sequences of words related to a topic until they come up with one that represents an actual sentence. Another question I have is how you could teach a computer to understand idioms. With the phrase, "hungry as a horse", the computer doesn't need to understand the exact quantity of a horse's consumption, but it may need to know more than that the phrase means "very, very hungry". It seems like in a language model, we might try to set certain phrases as equal, like "hungry as a horse" = "very, very hungry"; but in language this isn't always the case, since people consciously make a choice to say one or the other.
Edina Sarajlic
I enjoyed the Konane tournament very much, because it provided an opportunity to put the programs to their intended use. Observing the matches between two computers definitely improved my Konane playing skills. (This does not mean that I can beat a computer now.) I liked Peter's idea of giving evaluations of moves for the human player, because an inexperienced player (such as myself) can use these evaluations as a help in finding the best playing strategy.
Overall, I consider working on the Konane project beneficial in two respects: it has given me a thorough understanding of the Minimax search algorithm and significantly improved my programming skills. It has also introduced me to the dichotomy, which is in my opinion present in most of the AI topics we have covered in this class. Namely, the underlying ideas can have a very simple and elegant form, but their implementation is far from being such. This is very evident in reasoning and knowledge representation systems, which can be based on few simple rules, but require thousands of lines of code to implement.
Since we had only one class last week, I do not have any questions related to the reading material. My biggest concern at the moment is the final project, because I feel that we have much less time than we did for Konane. ( This is totally unrelated: I was just wondering if we are going to learn about semantic networks and their applications. The reason I am asking is that I have heard a lot about them - I do not even know if they are very important in KR part of AI )
Emily Sweeney-Samuelson
I read the handout about natural language understanding with great interest. Right now, that is one of the areas of CS that I am most interested in. I'd love a summer internship in that area, if anyone knows of any -- (ha, ha). But really, I am excited about this stuff. Does it seem like I say that about everything? Well, OK, so I get swept up in cool ideas. But this is no passing excitement of mine. I started thinking about natural language understanding (and generation) last summer, and it seems like a great way for me to go. I am a language/logic person, and I thought about majoring in linguistics, and this area incorporates all of that. It also seems that the future pursuit of natural language understanding will be exciting, popular, and productive. I hope so.
Of course, I should learn more about it before I jump to conclusions like the ones I just mentioned. It's still all just a feeling. But I enjoyed the handout. It laid out the components of the understanding process very well, so even the tricky concepts were clear. The diagrams were a great help to me, too, and the example sentences usually were, but I think they could have picked better sentences for a couple of them. It seemed like quite a broad introduction to many components of the problem, and made me think about many aspects of natural language understanding -- far beyond the technical implementation of such systems.
I enjoyed watching games in the Konane tournament on Monday. Person against person and program against program games were the most interesting to watch, in my opinion, but I've just never had an experience quite like that one before. It was fascinating to observe the differences between programs, and the lively atmosphere generated by a bunch of people in a room, all playing a board game.
Resolution in pred. calculus is good to sink one's teeth into. I like techniques such as that one, because the problems can often be solved in multiple ways. True, I don't always like a problem with that quality; I could live without the missionaries and cannibals problem. But resolution is satisfying and can be challenging, but not terribly (at least not yet.) It is just a little exercise, a good brain stretch. I hope I continue to enjoy it.
Natural language understanding is so exciting. Since I'm thinking of going into that area as a career, I'm devouring everything we hear about it in this class. Is there going to be a class on it anytime in the next couple of years?
Tim waring
After a fufilling thanksgiving break, i return to AI. I was able to show my relatives my konani program over break. I elicited an interesting response; most of them thought that it was pretty normal, what my program could do. (indeed they have had exposure to a number of game playing programs), and it seems that konani is a simple game and therefore a simple program. It is in many respects just that, but the search that it accomplishes is quite amazing, and at least part of many commercial games. This made me think about the number of games out there on the market that employ AI techniques. I also had a couple of realizations about standard everyday items, such as cameras and automated cars. My cousin, who has a masters in microbiology research talked to me about a new camera that tracks the position of your eye and focuses on whatever your eye is pointed at. He exhibited amazing distrust of the system saying "it's just a moke complicated machine, and the software programmers can't antipate everything, it's bound to make mistakes" Whereas that is true to some extent, he had now idea about the power of fuzzy logic, and AI devices working with out stimulus. Also my relatives were very amazed to here that cars are NOW able to drive them selves with little more than a modern desktop and a video camera.
It shocked me how AI techniques are in use all round us, and that the average american doesn't even know, or acknowledge that they are autonomous.
Sarah Waziruddin
Predicate calculus seems easier to understand than propositional calculus. The examples in class helped clarify and explain resolution and unification. Greens trick was also helpful and Im sure will be useful.
It was nice to see all the programs in the konane tournament and the many different ways that people chose to implement the program.
The Linguistics reading was a nice change from Nilsson and Moravec. I never knew this field was so extensive and would contained so much information. Computational linguistics seems to be the missing link to AI. We have studied reasoning and action and now communication completes the picture. However, a combination of these three things does not mean understanding but means that now, information can be manipulated, just like its done in Eliza. Understanding is not achieved. For example, if one has understanding then the meaning of ill-formed sentences can be guessed but without understanding, these sentences only lead to confusion.
Everything goes back to the very beginning of class, when we discussed the definition of intelligence. With these three tools, we have a seemingly intelligent machine depending on what your definition of intelligence is.
Leslie brought up a good point of body language. This is an important concept especially in terms of Eliza. Body language is extremely important for a therapist and partly determines their effectiveness. Generally, body language is extremely important in human-human interaction. I know software designers have thought about improving human-computer by working on user interface, etc. I was wondering if robotics has done any studies on body language. A robot in one of the videos smiled, frowned and blinked its eyes dependent on what the human interacting with it was doing and I was wondering if something had been done with body language to a larger extent in robotics.
Leslie Zavisca
Well, I hope everyone had a nice Thanksgiving holiday.
Resolution in first order predicate calculus makes sense to me. I've been able to follow all of the examples we've done in class. The only real confusion I've run into comes from keeping track of what two clauses were unified with which substitutions, but that's just an organizational aspect that's easily taken care of by being more careful and explicit about what you're writing.
The first session of the Konane tournament was pretty interesting. I only got to play one program version (Peter's) and the other two games I played were on physical boards with human players. Congrats to those whose programs won! What I would like to explore are the possible reasons that some of the programs were better than others. I don't know about everyone else, but even a quick examination of what sorts of specific implementation strategies (used by students in this class) are better than others would be valuable to me, especially since I don't have much programming experience.