INTELLIGENT SYSTEMS QUALIFIER

October 2004

 

All IS students must answer two of the following three questions:

 

Core 1

Consider the following problem for a robot.  The robot should search the environment for empty soda cans to collect, pick them up, and then drop them in a recycling container.  The robot carries a canister on its back in which it can store up to two cans.  It does not have a sensor to tell it if there is anything in the canister. It is provided the following behaviors (which may be considered ÒoperatorsÓ or ÒactionsÓ in the context of your favorite AI technique):

 

 

The robot uses continuous sensor inputs in each of the behaviors above, but it uses binary features to decide which behavior to use.  The robot is provided the following binary perceptual features that may be used to select behaviors:

 

 

Answer these questions in regard to the problem outlined above:

 

  1. Is it possible to use Q-learning to develop a policy for this task?  If so, describe how you would frame the RL problem for the learning algorithm.  If it is not possible, why not?  In either case be sure to fully explain your assumptions.

 

  1. Is it possible to sequence behaviors in a manner similar to an FSA to program a solution for this problem?  If so, describe an example solution.  If it is not possible, why not?  In either case be sure to fully explain your assumptions.

 

  1. Of the two representations for robot programs used above (policies and FSAs) which is more powerful and why?

 

* You can assume that if at-can is true, put-can-in-canister is 100% reliable
Core 2

It is said that a cat attempts to pass through a narrow space, such as the bars on a window, only if the width of space between adjacent bars is more than the size of his whiskers. Now consider a robot cat, which does not have whiskers. The robot cat thinks it can pass through the window bars if the coast is clear, e.g., there is no dog around.  Sometimes however the robot cat gets trapped between the bars as it attempts to pass through because the space between the bars is not large enough.

 

a. Write the following English-language sentences in first-order logic:

 

á      There is no dog around.

á      The window bar is passable.

á      If there is no dog around and the window bar is passable, then the cat will not be trapped.

 

b.    Now when the robot cat does get trapped between the bars, can it use logical deduction (by resolution, for example) to figure out that the bars in question were not passable after all? Show how (or why not)?

 

Core 3

Can analogy be applied to affordances? What are the arguments against applying analogy to affordances, in the traditional understanding of affordance perception? Are those arguments convincing, in your opinion? Feel free to use an example from any domain of research (e.g., behavior-based robotics).

 

 


If Perception is one of your specialization areas, answer two of the following three questions:

 

Perception 1

Vision is frequently proposed as a sensing technology for mobile robot localization and navigation. For example, a robot with a stereo head could simultaneously reconstruct its 3-D environment and estimate its motion. Suppose you are given the task of designing a vision-based robot navigation system for the Hall of Mirrors. This is a building with the property that every interior surface is covered with mirror glass. You can assume that all walls are planar and that any pair of walls are either parallel or orthogonal. Design a sensing and navigation strategy that would map the interior of the Hall of Mirrors. Describe both the sensing algorithms and navigation process in detail. Given an accurate map and a known starting point, how could you use vision sensing to follow a desired path to some destination? Provide a detailed algorithm description. You are allowed to configure the robot with any reasonable number of cameras (and light sources), but assume that no other sensors (e.g. dead reckoning, sonar-ranging, etc.) are available.

 

Perception 2

The Shell Game is a standard carnival game in which three containers of identical shape and color, such as cups or bowls, are placed upside-down on a tabletop. A small ball is placed under one of the cups by the dealer, who then shuffles the cups around, face down, while you watch*. After a certain number of moves in which cups are exchanged, shifted, and so forth, you are asked to guess which cup the ball is under.  In many real-life versions of this game, both money and cheating are involved. Here you are asked to design a computer vision system that could solve a fair version of this game. You can assume that you have a camera view looking down on the tabletop from above and that the dealer handles the cups by grasping them from above and sliding them around on the tabletop without lifting them. Although the dealer doesn't cheat, he/she is allowed to make distracting hand motions as well as ones that actually move the cups. Describe an approach to solving this game using the output of a single video camera. Characterize the core technical problem. Describe all of the necessary algorithms as well as your solution approach, and discuss possible failure modes of your system.

 

*You may wish to consult one of the many web versions of this game, such as http://www.free-web-games.com/games/shell.html, for an example of the procedure.

Perception 3

Suppose you are asked to track monkeys in a cage. The monkey cage has a video-wall on the back-wall portraying various jungle-scenes to put the moneys at ease. Various methodologies have been proposed for tracking targets with a changing shape and/or appearance. Three of them feature prominently in the perception reading list: Kass & Witkin's Active Contour Models, Cootes & Taylor's Active Shape Models, and the more recent Robust Online Appearance Models by Jepson et al.

 

a. Please contrast these three approaches within the context of the application

b.    which one you would use, and why?

c. How will your answer change if you know the video-wall will be turned off during these experiments?

 

 


If Machine Learning is one of your specialization areas, answer two of the following three questions:

 

Machine Learning 1

Naive Bayes (NB) classifiers have been shown to work surprisingly well for a wide range of classification problems. In some domains, substantial effort is required to improve upon their performance. Why is this surprising? How can you explain their success? Compare the following three classifier architectures: Naive Bayes, Neural Network, Support Vector Machine. Under what conditions would you expect each of them to dominate the others in performance? In your discussion, consider the following characteristics: number of features, amount of training data, presence of label noise, computational cost at run-time (testing phase). A NB classifier can be viewed as a statistical graphical model (specifically, a Bayesian network). From this perspective, one recent proposal to improve upon the performance of NB classifiers suggests adding additional edges between pairs of features to the model. Why would this be helpful? Comment on when this approach is likely to work well. Can you think of any reasonable criteria for deciding how to add edges, from the standpoint of both accuracy and computational cost?

 

Machine Learning 2

In J Yedidia, W. Freeman, Y Weiss (2003). "Understanding Belief Propagation and Its Generalizations", Yedidia et al. proposed an interesting scheme to do belief propagation in binary Markov random fields.

 

  1. briefly explain the scheme
  2. What is the big advantage with respect to belief propagation in normal DAG Bayes networks ?
  3. DAGs and MRF's are two ways to represent the same factored probability distribution. At least, every DAG can be represented as an MRF. Is the reverse also true ?
  4. What is the third way of representing factored distributions and how does that work ? Show an example of the Bayes network A -> B -> C <- D in both this third representation and MRFs, including the potentials.

 

Machine Learning 3

Sutton (1984), among others, has made a distinction between two kinds

of credit assignment problems: temporal and structural. What is the difference between the two?

 

Reinforcement learning is a learning technique that addresses the temporal credit assignment problem. Why is reinforcement learning suitable for temporal credit assignment but not for structural credit assignment?

 

Describe any AI method you know (or propose a new one) for addressing the structural credit assignment problem. Illustrate the method in a simple problem-solving domain, say tic-tac-toe.


 

If Robotics is one of your specialization areas, answer two of the following three questions:

 

Robotics 1

The DARPA Grand Challenge focuses on off-road navigation over desert terrain for long distances (around 150 miles) at fairly high speeds (25 mph or so).

 

(a)  What are two major technical problems in achieving a winning entry?

(b) The Army currently uses a variant of Albus' architecture, a derivative of NASREM referred to as 4D-RCS, for these sorts of applications. What is the basis by which navigation is conducted with this type of system in terms of sensing, planning, and acting?

(c)   Several of last yearÕs entries used behavior-based approaches. How would this differ from the approach described in (B)? What potential advantages/disadvantages would this yield over the 4D-RCS approach?

(d)  Describe what you would use for the perceptual capabilities for your entry (both sensors AND algorithms), which must navigate on poorly defined dirt and gravel roads, avoid obstacles and deal with barbed wire, brush and other vegetation. Assume you entry is a standard sport-utility vehicle (SUV).  Your design should incorporate existing algorithms (i.e. donÕt be concerned about creating an entirely new approach).

 

Robotics 2

The mapping overview paper by Thrun on the reading list provided an overview of recent SLAM approaches, including EM-based mapping and FastSLAM.

 

a. Please contrast both approaches in terms of what quantity they are optimizing, and what data is needed.

b.    (b) In a large environment, which one do you think will perform better, and why ?

c.  Is a FastSLAM approach feasible if instead of landmarks we use grid-based maps ? How ?

d.    how would you implement FastSLAM if instead of 2D landmark measurements we instead have a robot with an omni-directional camera-rig ?


 

Robotics 3

As robotics as a science continues to mature, ethical implications of our work become more and more relevant. Indeed Bill Joy, co-founder of Sun Microsystems, has gone on the record stating that robotics (as well as bio-engineering and nano-technology) may lead to the downfall of mankind.

 

(a)  Drawing from your own personal research in the area, explain the ethical consequences, both positive and negative, of the work you have already conducted or are currently conducting in this area.

 

(b) Given the apparent dangers associated with the field, why should you personally be granted license to become a robotics practitioner, especially at the Ph.D. level where you will articulate new research agendas? What are your own personal responsibilities (if any) to society in this regard?


 

If Planning and Search is one of your specialization areas, answer two of the following three questions

 

Planning and Search 1

Like many other AI systems, planners often model the problem of acting as continually repeating three steps: 1) sensing the world, 2) deciding on an action to take, and 3) executing that action. Also like many other AI systems, they often assume a synchronous world where each action takes a single time step and the system otherwise halts while an action is executing.  Imagine instead a more complicated model where multiple actions can be taken concurrently at each decision epoch. Assume further that each action is temporally extended; that is, each action may (perhaps stochastically) take more than one "step".  For example: while walking towards a car one might walk, reach for the keys, talk to a companion and avoid obstacles simultaneously.

 

First, describe several issues that arise in such an environment. Second, describe the strengths and weaknesses of your favorite modern planning techniques with respect to such an environment.  Finally, briefly discuss how you might extend some of those techniques and what the trade-offs would be in implementing your extensions.

 

Planning and Search 2

What is the relationship between planning and reinforcement learning?

In your discussion be sure to formally define planning, Markov Decision

Processes (MDPs) and partially observable Markov Decision Process

(POMDPs).  Further, explain both the technical difficulties that arise when formulating probabilistic planning problems as MDPs or POMDPs as well as the technical challenges of learning from data those representations (e.g., STRIPS and SAT encodings) that traditional planners use.

 

Planning and Search 3

STRIPS-type planning problems can be solved with search methods, such as depth-first search or variants of A*. Pick either GraphPlan or ASP. Sketch how the planner you chose uses search methods for planning. What are some of the limitations of the planner you chose? Speculate on promising research directions that might overcome these limitations (justify your answer).


If Knowledge Representation and Reasoning is one of your specialization areas, answer two of the following three questions:

 

Knowledge Representation and Reasoning 1

Nebel and Koelher (1995) (see citation below) provide complexity results for case-based planning that are quite pessimistic. In particular, by expressing the plan generation and the plan modification problems in a STRIPS-like propositional language, they are able to show that the worst-case time complexity of plan modification in general is no better than that of plan generation. This, they claim, is true even if plan modification is conservative, i.e., it reuses as much of the old plan as possible and makes minimal changes for achieving the new goal (which is what case-based planners typically do). One reason for this, they say, is that the complexity of determining the maximal reusable part of the plan itself is in the same class as that of generating the plan.

 

This of course flies against the intuitions, hypotheses and experiences of case-based planning. How may one reconcile the theoretical results of Nebel and Koelher with the experimental evidence for case-based planning?

 

[citation: http://www.informatik.uni-freiburg.de/~koehler/papiere/planning.html ;  to save time, look at the short IJCAI paper rather than the long AIJ one; also focus on the introduction, discussion and concluding sections, not on the proofs of the various theorems.]

 

Knowledge Representation and Reasoning 2

 

a. What are the differences between logic-based and justification-based truth-maintenance systems (TMS)?

b.    Pick any problem-solving domain, and determine whether a justification- or logic-based TMS would be most suitable. Justify your answer.


  

Knowledge Representation and Reasoning 3

Model-based autonomy (Williams and Nayak 1996; Muscettola, Nayak,

Pell and Williams 1988) is a recent paradigm for designing autonomous agents. The key idea is that if, in addition to a continuous model, a discrete model that captures the causal structure of a hardware system is available, then model-based diagnosis can be used to identify causes for malfunctions in the hardware and to recover/reconfigure/repair the system.  Since model-based autonomous systems have been deployed aboard Cassini, a recently launched spacecraft for exploring the moons of Jupiter, they make an AI success story.

 

However, an important limitation of of model-based autonomous systems is that they are applicable only to immobile agents; in case of Cassini, model-based autonomy is limited to the immobile engines of the spacecraft, not to the mobile spacecraft as a whole.

 

a. Why is model-based autonomy as practiced by Williams and Nayak limited to immobile agents? That is, what is the source of the limitation?

b.    What additional issues arise in realizing model-based autonomy for mobile agents that can explore an environment, make decisions, form plans, etc?

c. How might the paradigm of model-based autonomy be adapted to address autonomy in mobile agents?