Intelligent Systems Ph.D. Qualifying Examination - Spring 2001

1. Opportunistic planning/reacting is an understudied area in mobile robotics, and even more so in multi-robot teams.
1. What does it mean to be opportunistic in mobile robotics? Give an example that is particularly relevant for multirobot team applications.


2. A few architectural approaches have explored this, notably Alliance and Atlantis. Briefly describe how one of these (or another relevant architecture) attempts to permit flexibility through opportunism in mobile robotic systemss.


3. Explain the role that communication can play in introducing opportunism to robot teams. How can decisions be made regarding what is important to convey among team members? Distinguish between those things that are of global value to the robot society versus those things that may be pertinent to one or just a few team members. Share your ideas on this relationship between communication content and the effective exploitation of opportunities in a robot team. (Remember communication can go in both directions)
2. 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.
3. 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 strategy or method to use next?" Enumerate and briefly explain the criteria an agent may use to select a method from the various methods available to it.
4. 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.
5. Assume that one has to learn classifiers for two related but different inductive learning tasks with similar features, say classifying a picture as showing a one or two lane road and finding the location of the left edge of the road. One could do this by solving two different inductive learning tasks, for example by training two different neural networks. However, some researchers argue that there are advantages to solving both inductive learning tasks together, for example by training only one neural network whose outputs are the union of the outputs of the two neural networks (that is, it has outputs for classifying the kind of road and output for determining where the left edge of the road is). What are the advantage and disadvantages of this approach?
6. Why are planning methods for planning in deterministic domains often very different from planning methods for planning in probabilistic domains? Explain at least three different approaches for how one could apply (or how researchers have applied) methods for planning in deterministic domains to planning in nondeterministic domains. What are the main obstacles that one needs to overcome for this research direction to succeed?
7. Some researchers study meta-reasoning architectures, including how much a system should reason about its reasoning process. Sketch at least three different approaches for determining how much to reason. What are the advantages and disadvantages of these approaches over approaches that do not reason about how much to reason? Why aren't more systems out there that use principled approaches for determining how much to reason?
8. Human control of distributed robots has become a topic of interest in the robotics community. Many researchers are now considering wearable computers for this task. What are the issues most important in determining whether a wearable computer is appropriate for the job? Be specific by discussing particular applications and the advantages and disadvantages of the interface.

   How might we use this problem domain to research artificial intelligence? What might we learn?

9. Darrell and Pentland wrote a paper called "Space-Time Gestures" where they showed a computer vision based gesture recognition system which recognized a few gestures in a continuous stream. Their approach was to take a training set of frames of the various gestures, run principle component analysis on them, and store some set of these principle components. Then, for each trained gesture, each frame making up that gesture was projected on to the principle components and the projections saved as a template for recognizing the gesture later. In testing, again each frame was projected on to the already trained principle components. Dynamic time warping was used to compare the progression of projections in the test gesture to those learned in training.

   However, many sign language gesture recognition systems have concentrated on using hidden Markov models. Explain the advantages, if any, of hidden Markov models over dynamic time warping. Feel free to use diagrams or particular problem domains to help explain your points.

10. Explain how Kalman filtering can be used to make motion capture less noisy. Where could predictors for human motion come from?
11. Explain 3 methods to solve redundant inverse kinematics.
12. Humanoid robots typically have a visual system on a moving head. Explain 2 methods to touch a visual target with a hand in the presence of head and body motion.
13. A robot has to navigate a crowded factory; most of the obstacles are humans so running into them is a bad idea. Because of the density, the robot is going to use a stereo vision system to detect how to get around the humans. What would be a good stereo algorithm to use in such an environment and why? How would it compare to, say, a sonar based system?