CS 372 (Artificial Intelligence) Fall 2004
Week12 (Due 11/22) Reactions & Thoughts
Ioana Butoi
Andrew Cantino
It is interesting that FOPC is so good at some things, but has such trouble with things like exceptions, which humans have no trouble with at all. I also find it interesting, as was mentioned in class once or twice, that humans can easily and instantly tell you whether they know something or not, when a computer would have to search through all of it’s knowledge to find the answer. I think this highlights the fact that humans and computers work in very different ways. We may both be able to do logic, but on the inside, we are very different creatures.Working on the konane project has been interesting. I’ve made my graphs, and I see an interesting ‘ripple’ like effect in the graph of avg. branches vs. search depth. I think this has to do with whether we are ending our search on a min or max node, but I’m not entirely sure how this works. I’ll bring it up in class.
Audrey Flattes
Programming commonsense knowledge is really interesting when I think about it and at the same time strange. The thought of programming common sense is so funny. It just reminds me that when dealing with machines they are given nothing and we must teach them everything. Then my train of thought leads me back to the question of AI and wether in the end it will really matter. Even if we can program systems to act and simulate thinking where will that take us. I'm totally going off on a tangent here but, what is the end goal of it all? Are we trying to create agents that will serve to help us with our lives making them easier for us or are we trying to create agents that will live and be equal to us in society(which I don't think will ever happen)?Christina Florio
While reading chapter 18 my head kept filling with questions. Such as, could commonsense knowledge be learned? For instance, if you told it once what to do and it was able to make the connection between the action and its result in order to use both later. Or, maybe the robot could test different actions on something and then it could decide which action was best or if it needed to find a new action. While basic information would be required, the robot would be able to infer on its own, therefore, expand its knowledge. It seems like if a robot were able to learn as it was working, much like humans, then it would be able to gain the necessary commonsense.
Kathleen Maffei
While reading chapter 18 about common sense knowledge, I was reminded of learning robots that I’d read and heard about. It seems to me that if you want to imitate the set of “common sense” knowledge that humans generally hold, then you might create that database the same way that humans do: by starting the program out as a “baby” and feeding it examples. Since children also learn through play, it would be helpful if the agent could interact with its environment. But children learn also by observation, and so information could be fed to the agent in the form of examples of everyday situations – text or video or pictures, depending on what’s possible. It seems to me that this database, once created, could be transferred to any other agent which uses the same logic encoding.
I definitely see where so-called common sense knowledge would be useful. Of course, a brute-force expert system might sufficiently make up for the lack of common sense, but I think the flexibility of the logic systems that would be needed – default logic, time representations, semantic structures – would contribute to making the agent think more like a human.
Sara McCullough
I am very intrigued by the commonsense knowledge problem in AI. I am surprised that there has not been much discussion of it up to this point. It seems like the hurdle would be the most pressing problems to solve in AI. I agree with Nilsson in that commonsense knowledge is what is needed in order to make AI systems more useful. I think that if we could give computers the ability to learn and reason based on a commonsense knowledge base, the field of AI would enjoy a massive step forward. I also really like the questions that the commonsense knowledge problem raises. For instance, the question of how to represent time is something that is very difficult for me to think about. Time is difficult to fathom. Yet using and understanding time is most definitely commonsense knowledge. The question of how to represent it is something I am definitely interested in studying more. I think that an expert system with a large commonsense knowledge base would be amazingly successful in learning, reasoning and deducing things about the world and environment.Ben Root
I really enjoyed hearing about everybody’s final projects in class on Thursday. I think that the neural net static evaluation function for Konane is an awesome idea because my minimaxAB function is just racking up staticEvals. The idea of having a constant time static eval sounds SWEET.
I’m not sure whether the static evals are the bottleneck but my minimaxAB takes it’s sweeet time to do 3-ply depth searching. I may try to order the nodes with the ones having the best 1 level deep static eval getting searched first which may cause a speed increase, but right now it is really boring to play against the computer when it searches 3-ply because it won’t make conversation while it thinks. Maybe that is the solution, I will make the computer talk to me:
“I’m evaluating the move from 5,3 -> 7,3 …it’s minimum static eval is: 5. I’m going to check out the rest of the moves before deciding. Now, I’m evaluating the move from 6,4 -> 2,4… etc.”
Sandeep Singh
AI agents often remind me of idiot savants. On the one hand, while an ‘intelligent’ computer can crunch massive numbers and search large trees with great ease, they nevertheless lack common sense knowledge. As mentioned in Chapter 18, it seems like the only real way to create a knowledge base of ‘commonsense’ is to manually enter each assertion. We saw with our human robot the difficulty faced to the agent when posed with the simple task of placing one box on top of another. For humans, such a process doesn’t appear to be so specific—moreover, it’s commonsense and eventually becomes mechanical. While some of the things we consider commonsense were actually learned through example and trial and error, other things appear to be genetic and innately known. For example, a hungry baby will cry when s/he needs food. However, the baby will soon understand that whenever s/he cries, they will (hopefully) receive immediate attention from the parent. Through positive and negative reinforcement, a child will learn what actions will produce positive and negative reactions. If a child wants to play with a toy that is being used by a playmate, commonsense tells the kid to grab the toy. When the child is punished for doing so, s/he will (or should) begin to realize that doing such an action is wrong…which then becomes commonsense in itself (the Golden Rule). This system of stimuli and responses is how humans grow to learn. Personally, I don’t think it’s realistic for an AI agent to come up with a commonsense knowledge base on its own. We really can’t slap an agent on the hand when it does something wrong or give it a cookie when it does something right…or can we? However, given that it does have an initial knowledge base, I think it is possible for an agent to deduce new commonsense assertions through experience. This brings us back to square one…how can an agent build up a ‘commonsense’ knowledge base without a human manually dictating the assertions?
Darby Thompson
Commonsense Knowledge seems to be a really huge hurdle for AI to overcome. How is it that we are born with some degree of commonsense? Is there a distinction between commonsense and instincts? In reading this chapter I have been reminded of the documentary we watched at the beginning of the semester. There was one experiment where a duckling was shown a shadow of a duck flying above, and then a shadow of a predator flying above. The duckling had an instinct / commonsense to realize that the first was of no threat, and the second was very scary. So how can we ‘give’ an agent that which we were given – but with no knowledge of how we received it, what its capacity really is, and how we store all this information, organize it and use it in the correct circumstances? Do we even need to worry about giving commonsense to agents? Well… Nilsson seems to think so, and his argument is pretty convincing to me – commonsense is needed in order to make expert systems more useful, so that they can adapt to situations not necessarily directly in their field of expertise. What use is a household robot if you have to answer questions constantly? What we need is a robot that is good at problem solving… an agent that can come up with a solution without being told the answer. If it doesn’t know where the plates are kept, it should figure out the most likely locations and then explore and see if it can find the plates.
