Please use this exam as a opportunity to demonstrate your knowledge in the various IS areas. Several of these questions can be answered from a variety of perspectives or "area points of view". Please be explicit about how you interpret the question, what assumptions you make, and what representations and algorithms you use.

1. The field of AI has often used the cognition of people to inform itself on how computers might perform intelligent tasks. Much of the work on goal-directed reasoning is based on human cognition, and even expert systems take their inspiration from the reasoning of human experts. More recently, AI practitioners have begun to think about how the understanding of cognition we gain from our computer models could inform human reasoning. Learning by Design and Goal-Based Scenarios, both based on case-based reasoning's model of cognition, for example, focus on helping learners learn skills and content knowledge from experience -- orchestrating the environment so that they reflect in ways that promote productive interpretation of their experiences. Case-based reasoning tells us about the kinds of content that make cases especially usable and about the kinds of interpretation that make them particularly accessible, and LBD and GBS propose ways of making sure, in the classroom, that students reason about their experiences in ways that promote creating those kinds of representations and indexes.

What you need to do in this question is to consider the implications of structure-function-behavior theory (SBF) for promoting human learning. SBF theory helps us understand the ways in which computers can productively reason about systems, models, and devices. (a) What, specifically, does SBF posit as a model of such cognition, and (b) what does that cognitive model suggest about the kinds of reasoning we might want learners to do as they are reasoning about models (in science, economics, engineering, ...) or as they are learning about systems or devices? (Ken Forbus, by the way, is working on a variation of this challenge.)

2. There is a tension (both historically and intellectually) regarding the role of knowledge in behavior-based systems. Dan Dennett has characterized knowledge as taking three different representational forms: explicit, implicit, and tacit. There may be other useful taxonomies as well, but use the different qualities of knowledge to answer the questions below.

1. The claim that a truly reactive architecture is completely free of knowledge representation has served as the hallmark characteristic of nouvelle AI. Is it really true though? How according to Dennett's (or some other) taxonomy can knowledge be viewed as being represented even within a system such as the subsumption architecture? Or is it really not at all?
2. Payton, Borenstein and others have added the use of knowledge representations that still permit the use of a behavior-based reactive system design, but still allow for explicit knowledge usage. How can this be? Why bother? What advantages does it afford? What consequences can occur if poorly managed?
3. Describe a prototypical design for a behavior-based robot architecture that can be used for navigation in a deep underground coal mine. Assume that new tunnels are created on a daily basis, but the old ones persist over time. How could you introduce this ever expanding map knowledge dynamically, incorporate it into planning processes, and yet still deal with the overall dynamic nature of the mine (changing topology, miners, coal cars, hazards, etc.)? Be very specific regarding specific algorithms, knowledge representations, etc. to give a complete description of the capabilities of your system when compared to a purely reactive one.

3. You have to develop a controller for a (physical) pinball machine. The controller senses the position of the ball 10 times a second and can make both flippers flip up at the same time (in which case they flip down immediately afterwards). Initially, the controller knows only the layout of the playing field. Its goal is to play as long as possible. The game is over once it has lost the ball 10,000 times.

You would like your controller to use reinforcement learning techniques. Explain how your controller would use them. (Be as specific as possible.) Then explain how your controller addresses the following issues: a) that it is not able to predict the movement of the ball with certainty, b) the time constraints, c) that it has to learn while performing (in other words, that it has to explore while exploiting), and d) the time horizon of 10,000 balls.

4. Explain the advantages and disadvantages of planning with abstraction hierarchies. How could one integrate abstraction hierarchies into search-based planners such as ASP or LRTA*? What problems show up and how could they be solved?

5.

1. Explain how you would implement a call routing system in which callers could use natural speech to navigate an automated phone answering system. By "natural speech" we mean utterances one might say to another person.
2. Explain how you could use techniques from information retrieval, such as a vector space representation of text, to correctly handle caller's requests.
3. Explain how you would get the system to ask for clarifying information from the callers, both for your answer to A) and your answer to B).

6. Explain how the ability to generate a movement can assist in perceiving it (the motor theory of perception). How do internal models help a robot know where it is, or know the trajectory of ball thrown through the air. What is an alternative to this approach to perception?

7. Explain how neural nets can be used to implement symbolic reasoning. In the process of answering this question, please explain what the differences are between symbolic and numeric approaches.

8. A well-known issue in intelligent systems is their need to interface programs with some external environment. Since the external environment acts in real-time whereas programs do not (necessarily), when intelligent systems programs are run, special care must be taken to coordinate their execution. Rate monotonic scheduling and deadline-based scheduling are two methods commonly used for IS program execution. First, distinguish RM from EDF (earliest deadline first) scheduling, by defining both. Second, explain how both may be implemented. Third, explain the following statement: "EDF scheduling is dynamically optimal", by commenting on what optimality means in a dynamic system.

9. Simulations of ant colony behavior have been proposed as the basis for optimization algorithms. What optimization approaches can we take from animal behavior, and what problems are they good for?

10. Why would humanoid robots be worth doing research on? What intellectual issues will they help us solve? Why can't we research these issues on industrial robots? What are the key research issues for humanoid robots?

11. Why do the N-gram language models used in speech recognition work so well (as opposed to grammars)? How are they used? How are they learned? What will replace them?

12. Most knowledge systems use propositonal representations. Let us consider, in contrast, a knowledge system with a depictive representation, for example, a medical knowledge base consisting of radiologic images. The problem of accessing knowledge relevant to a reasoning goal, of course, is as important in an imagistic knowledge system as in any other. Describe a scheme for organizing an imagistic knowledge base so that a reasoner can "zoom-in" and "zoom-out" of an image and access it at the grainularity appropriate to the reasoning goal.

13. Let us consider two different kinds of "mental models:" (i) "scripts" developed in the context of AI theories of natural language understanding, and (ii) "functional representations" developed in the context of AI theories of problem solving and design. Compare and contrast the representational and organizational commitments the two types of models make. Argue that the representational commitments made by each are appropriate to the task they address. Critique the two models from the perspectives of their own and each other's tasks.