CS372: Artificial Intelligence (Fall 1998)

Week 5: Responses

Readings

• The Robot: Corral Exiter, Obstacle Avoider (Subsumption-Based)
• Chapter 7, from Nilsson's text.
• Chapter 2 from Moravec's Mind Children

I was also wondering if there had been more progress with creating a working three-dimensional circuit using a diamond? I had never heard of such a thing before but it is an interesting idea.
The other interesting and far-fetched-sounding idea was the possibility of tiny robot arms which would be able to grab individual molecules (nanotechnology). Something like that would seem to me to be very far off in the distance at this point and I wonder if that would ever be achievable. I wonder why we would need a computer to be able to do this? Or why we would need a computer with capabilities of "10^30 times the power of the human mind"...? I cannot even conceive of what we might use them for. Sounds scary, though.

Finally, lab this week seems to be taking a long time. There were many things to consider when trying to get our robot to do exactly what we wanted it to do. I suppose it takes practice.

Well, since I'm trying not to think about the test right now, I'll write about the problems of building a robot which uses subsumption architecture to escape the corral and find the light, so to speak.

At first, the matter seemed simple. The basic "cruise" command could go forward until it bumped into a wall, or detected a wall with its infrared sensor; at that point, "cruise" behavior could be subsumed by "wall-following" behavior, and the robot would travel around the perimeter of the corral until it came in contact with light. at that point, the light falling on the sensors would trigger the highest-priority behavior, "light-following", to kick in.

In practice, though, subsumption architecture didn't work as planned. given an obstacle in front of the light, the robot started traveling around the obstacle, performing excellent "wall-following" behavior. Unfortunately, this action was never subsumed by light-following behavior, because the light-sensors never noticed light to the side of the robot, even after fiddling around with the physical placement of the sensors on the robot.

We are trying a different approach now, slower, but perhaps it will work in the long run. By having the robot systematically cover every space in the corral as its "cruise" command, the robot will travel back and forth across the space, slowly moving up the corral until it encounters light. This method may take longer in the end, but perhaps ith will work, because the robot will eventually turn to face the light.

The problem we experience with our first program, though, brings up an interesting dilemma. Even if you have the greatest program in the world, it might be completely useless because of the structure of the actual robot. I have noticed that in our discussion of artificial intelligence so far, there has been very little talk of the robot "body", with the exception of Moravec. I get the impression (I may be wrong) that the "mind" of the robot, as in programming, etc, is considered almost entirely separate from the "body", or structure, of the machine. This seems to me to be a mistake, one based on the traditional separation of mind and body when considering human behavior. When psychologists began thinking of the\ mind--or brain--as just another part of the body, affected by the stomach, the heart, the liver, etc, it revolutionized the way the mind was perceived.

Perhaps a partial return to cybernetics would allow AI researchers to better unite robot structure and programming. Perhaps programs should be built around the structural design of a robot, rather than the structural design being built around the program.

A maximally "surprising" message, he states, yields the higher information content (according to information theory). And he cites as an example: "Here's my cat. It has fur. It has claws..." and "Here's my at. It wears a hat. It totes a gun.." saying that the first series is "boring" and the second is "better." Better by what standards? If you put two machines in front of me and one produced the first series of phrases and the other the second -- I'd say the one who left out the gun would be the smarter. Well, maybe not smarter as Moravec uses the example in conjunction w/doing "useful work." But then again, a more "correct" answer would be more useful than a bizarre one. It's more likely for a cat to have claws then a gun. So maybe it was just a bad example. Or maybe its a good example, but used in the wrong context. There are times when a "gain" in information is useful -- e.g. genetic algorithms, cluster analyis, and other machine learning techniques -- where a process that yields more change between two states is better than one that only yields little information gain. There are also times when a such a gain could be bad -- too much change could yield a bizzare result. That leaves me with a question -- how is he using the term in the book and how exactly does information gain potential relate to computational power (i.e. what he is trying to show) ?

I see that many actions are limited by reactive agents because of the memory that would have to be involved to effectively react in all situations. And I understand that computations might be more efficient. But it still seems to come down to the same problem: unless the world state is perfectly static, how could one forsee all the needed equations? As Nelson states, these agents are ideals, and "do not hold in most real-world applications". Are its limitations really much different from a reactive agent that needs a superhuman programmer or a small environment to function?

I think it's interesting that Moravec began chapter 2 on "evolution" (of sorts) with what amounted to an excuse as to why AI progress had slowed down: the culprits being the electronic toys we love, like email, graphics, and other "luxuries". Once again human intelligence is the meter by which we measure progress. Instead of focusing on certain jobs a machine might easily perform, perhaps something more efficiently performed than in human hands, he concentrates again on achieving human intelligence, in what he calls the "joyride of human equivalence". It still strikes me that AI should have so many diverse, and sometimes conflicting, goals. I thought it was interesting that he launched into a long biological description of the neuron and the human retina, only to end quickly by saying that the TV next to him had the same capabilities. It was almost as if he needed a biological justification for what pushes AI; perhaps this is how he feels the reader needs to justify, or at least understand it. Can't I assume that in the past 10 years that computers have been made to compute the 10 frames a second he was missing?

And then there was his jump to calculations for brain activity.... I wonder if we can feasibly calculate brain activity as a function of the neurons in the retina. After our class discussion about the Sony cat, I'm not sure that we are 10 years off his prediction of fully intelligent machines. I was thinking about information theory as well. How can you tell if unfamiliar information is "surprising"? It was interesting to see the "century of computing" graph as well. But I wonder if the trend will continue; it seems that the closer we get to achieving AI, the more complicated the problem becomes, the more the process might slow down, or become blocked. A good part of the century was spent in mechanical progress. Now that the chemists, biologists, and physisists are involved, it might be that the complicated process might slow down.

I've enjoyed class discussions, especially when we learn about which new machines are on the market, or what is used in competitions. Judging from my friends' reactions to our robot, I think we could try to sell them in the bi-co.

David Rothstein

In reading Chapter 2 of Moravec's book, the first thought that comes to mind is that his predictions are probably too optimistic. As we've discussed in class, his prediction of a household robot within 10 years of 1988 did not come true, though the technology probably exists (in various separate components) to make such a robot. Similarly, it is doubtful that his prediction of a human-like robot within forty years will come true. What Moravec does not take into account is that although the technology to create a certain type of robot may exist at a certain point in time, the time and effort required to bring together and make use of this technology is often not practically feasible. We may be able to construct practical-sized robots within 40 years that have the computing power of the human mind, but it would take an enormous amount of time to program such a robot to act like a human. On a much smaller scale, my group ran into this issue in preparing for the upcoming robot lab this week. Our robot was running into the problem that it often got stuck in corners. Thinking of and being able to verbally articulate a routine that would get the robot out of the corner was extremely easy -- there are many instructions a robot could follow to initiate this escape. But going through our program and actually making the changes that would implement this routine was a much more difficult task. In short (as the saying goes), it is easier said than done.

Conversely, though, what really interests me about Moravec's book is his assertion that creating a human-like robot is technologically feasible. Even if he has the time scale way off, the fact that such a robot may be within the realm of the technology that we are beginning to see today means that such a robot eventually CAN be built, and therefore it probably will. Even if it takes as much as a thousand years rather than 40, this is still an exciting thought, and one that probably serves as a motivational tool for many AI researchers.

n this weeks lecture we have started learning about reactive agents that are capable of planning. I was very excited to learn that it is possible to predict the consequences of an agents actions many steps ahead, and then to pick the steps that would ultimately bring the agent to some goal state. Since we spent a lot of time working on our lego bug, Princess Leia, I could not help but wander if it would be possible to implement some form of planning in her behavior. Nillson lists a couple of prerequisites that an agent needs to have in order to be capable of planning: the model of the world and a model of the effects that its actions have on that world. Leia is at this point only a stimulus response agent and she acts based on immediate sensory inputs only, because she does not have an internal state where she could store the representation of the current world state. Also, the only way she can affect her environment is by changing her position in it, so not having a representation of the environment excludes the possibility of planning.

In this chapter Nilsson talks only about explicit state spaces, where a simple algorithm can be used to find a sequence of actions that would lead from the initial to the goal state in the state space. However, as Nilsson mentions, explicit state spaces are rare in real situations. I am very interested to learn about planning algorithms used in situations where a) all possible states of the world cannot be foreseen, b)there is not an exact correlation between an action and its effect, and c) where the goal state is a dynamic quantity, say a function of the current state. I guess it is possible to deal with the last case by projecting at every other step, checking the current world state and computing a new path to the goal state. The second complication (b) can probably be remedied by representing different possible effects in a state space graph with larger branching factors. I guess all of these situations reduce to one problem, the one of creating and searching large state-space trees. This must happen quite open in real-life applications, so I hope we discuss it in class.

In chapter 7, Nilsson introduces the idea of an agent reacting to situations that it has not been programmed to react to. For this, an agent must be given a model of its world and a model of its world after being affected by some of its actions. State-space graphs are used to implement this idea. Schemas are used in these graphs and they model the affects of an agents actions on its world. It seems as if the robots job is to find the optimal path from its initial state to its goal state. There are two kinds of state-spaces, iconic based and feature based. Moravec focused on the physical aspects of implementing intelligent behavior in chapter 2. He pointed out that there weren't good programs for artificial intelligence and there was (is?) also a lack of powerful machines. Moravec approximated that the brain performs 10^13 calculations per second. He stated this speed was faster than the speed that computers could operate at when he wrote the book. I wonder, is this still true? Can current computers perform 10^13 calculations per second? If they can, has this greatly aided progress in AI? Moravec also brings up the issue of memory and again, I ask if computers today have enough memory to match human memory and if they do, does this aid AI? I am disturbed by the idea of developing new proteins that would help form new technology. This is messing with creation and very similar to cloning. I agree with Maralee with her views on God and I ask, what are we REALLY trying to achieve by having thinking machines?