INTELLIGENT SYSTEMS QUALIFIER
FALL 2000

Please answer any seven (7) of the following thirteen (13) questions.

1. In pattern recognition, a distinction is often made between methods that are good at reconstructing a set of data versus discriminating between different classes of those data. Give examples of situations where particular techniques may be good at one but not necessarily good at the other (for example, recognition systems that are not good at reconstruction of the test data). For each example, propose a complementary system that may handle the case not handled by the technique described (either discrimination or reconstruction).

2. Why is subpixelization in computer vision often important when registering 3D graphics in augmented reality?

3. Why is on-body perception of the user with a wearable computer more difficult than self-perception on a robot (as one example, in determining location)? How might this situation be improved or even turned into an advantage?

4. A typical approach to gesture recognition is to segment the incoming data stream into some fundamental unit (like the phoneme in speech) and then perform recognition on each segmented unit. Why might this be a poor choice? Give an example. Why is context and language modeling so important in modern speech, handwriting, and gesture recognition? How might you apply the lessons from these systems to recognizing and modeling more generalized human behavior?

5. Compare Bayesian networks and influence diagrams against totally and partially observable Markov decision process models. What are the strengths and weaknesses of these tools from the point of view of 1) knowledge representation, 2) knowledge acquisition, and 3) reasoning? Speculate on how one can combine the strengths of these tools and describe the kind of research results that are necessary to achieve this goal.

6. Assume that you are on NASA's team to develop better AI planners for controlling autonomous spacecraft. The planner is always in charge of the spacecraft and has to plan all spacecraft maneuvers (including emergency maneuvers). Describe the strengths and weaknesses of STRIPS-based partial-order planners in this context, and suggest five promising research directions for how to make this and other artificial intelligence planning technology better suited for the task.

7. The notion of marsupial robotics is beginning to take root in the robotics community. This is where one robot serves as a launch vehicle, recharging station, communications link, etc. for a smaller more agile one. This raises whole series of questions regarding vehicle coordination.
   1. Give a task-environment where marsupial robotics might be put to good use. Why is it better to have strong dependencies between robots in this case, rather than simply having multiple identical robots?

   2. It might seem that a hierarchical or deliberative software architecture is best suited for this type of system at first glance due to the parent-child relationships. Why might this not be a good idea particularly in dynamic environments?

   3. Now suppose you had to distinguish your own parent robot from a host of other parents (the problem of kin recognition) in order to dock with it for recovery and transportation to another site. How might you accomplish the recognition and docking task? Describe in detail both the perception and motor controls strategies that you would use.

8. Machine learning methods could be categorized as reflective (involving some strategic monitoring and planning) or non-reflective.
   1. Provide a more technical definition (your own) of what it might mean for a machine learning method to be reflective or non-reflective.

   2. Identify at least two examples for each category, and justify why each belongs where you've put it.

   3. Identify the strengths, weaknesses, and maximum potential of each category of approaches, and the difficulties in achieving those potentials. Back up your arguments with plenty of concrete examples.

9. Consider the following common AI tasks: interpreting the meaning of a visual image (or a sequence of images), interpreting the meaning of a sentence in a natural language (or a sequence of sentences), and interpreting the meaning of a behavior (or a sequence of behaviors) of a system, an agent or an environment. The meaning of "meaning" is an important issue in all of these tasks. Enumerate and briefly explain the various meanings of "meaning" in AI.

10. Control of processing in AI generally refers to the question "what to do next?" One manifestation of the control issue pertains to the question "what method to use next?" For example, a system (or an agent) may know both the method of genetic algorithms and the method of backpropagation in feedforward neural networks, and, in any given situation, may need to decide which of the two methods to use. Enumerate and briefly explain the criteria a system may use to select a method from the various methods available to it.

11. Before random-dot stereograms (RDS) [same random noise pattern in both eyes except one region moved to produce an effect in depth], it was presumed that objects in one image were matched with those in another to achieve the stereo effect. The claimed power of RDS was to show that the human system didn't need blobs or any other discernible structure to do stereo. Argue both sides. That is, why would matching blob-like structures be better than pixels, and visa-versa? Also, do RDS really argue for no matching of structured elements (think multiscale)?

12. Enumerate the basis tenets of Brooks's subsumption architecture in behavior-based robotics. Discuss how these tenets might apply to interface design on a wearable computer.

13. Context and relevance are important notions in retrieving information from the internet. One way of capturing the context is to overlay a semantic network on interrelated webpages and use spreading activation to access webpages relevant to a query. Describe how this scheme may work in practice.