INTELLIGENT SYSTEMS QUALIFIER
Fall 2009
Each IS student has two specialty areas. Answer 2 of the 3 questions
in each of your specialty area as well as 2 of the 3 Core questions
below.
=============================================================================
All IS students must answer 2 of the 3 Core questions:
Core AI #1
Russell & Norvig's popular AI textbook says that their goal is to "design
agents that act rationally". They then present three possible meanings of
"rationality":
1) Perfect rationality: A perfectly rational agent acts at every instant in
such a way as to maximize its expected utility, given the information it has
acquired from the environment.
2) Calculative rationality: A calculatively rational agent eventually returns
what would have been the rational choice at the beginning of its execution.
3) Bounded rationality: A bounded rational agent behaves as well as possible
given its computational resources.
Part A: Which of these do you think should be the goal of Artificial
Intelligence as a scientific field of study? Justify your answer.
Part B: Which of these do you think should be the goal of Artificial
Intelligence as an engineering discipline? Justify your answer.
Part C: Which of these has been (explicitly or implicitly) has been the goal
of your own research thus far? Explain.
Core AI #3:
In the 1990s, there was a surge of interest in "Autonomous Agents" in HCI.
The idea is that "intelligent" software agents would observe your repeated
actions on the computer (e.g. sorting photos), determine what you are doing
(e.g. moving all the even numbered photos in one directory and all the odd
numbered photos in another directory), and offer to automate the process.
More recently, in ubiquitous computing, there have been attempts at
"contextually aware" systems. One example is Scott Hudson's work at CMU
where they tried to make a system that could determine how "interruptible" a
person was and delay alerts from the computer if the person was deemed to be
too engaged in a task.
The canonical example of a contextually aware system is a mobile phone that
knows when to ring audibly and when to just vibrate. Much machine learning
and AI tricks have been devoted to this subject.
A) Argue why this last problem is an example of an "AI-complete" problem.
Some researchers have argued that the problem is not really "AI-complete" but
"mind-reading complete." In other words, the problem is not really about
making a human level intelligence that can determine when the phone should
ring versus vibrate, because even another human can not predict the right
answer all the time. The right answer really depends on the user's
unobservable mental state, making the problem harder than "AI-complete."
B) Give two other examples of problems (not using the 3 above) that seem at
first "AI-complete" but are really "mind-reading complete."
C) Describe how you might use AI-related techniques to do a fair job at these
problems even though they are "mind-reading complete."
Core AI #3
The faculty often joke that AI consists of clever ways of cheating "NP-hard"
problems.
A) Give example domains/techniques where this viewpoint seems true.
B) Give example domains/techniques where this viewpoint seems erroneous.
========================================================================
If one of your two areas is Knowledge Representation and Reasoning, answer 2
of the 3 questions below:
KRR #1
In narrative psychology, inference is considered an important part of
comprehension. A reader reads a sequence of events (sentences describing how
character actions change the state of the fictional world). It is assumed
that inference is used to determine whether one event is a causal antecedent
of another, that is, the former event is required for the latter event to be
possible. Suppose you have a system that tries to determine what events can
be taken out of a story such that the story is still comprehensible. The
system uses a guess-and-check method. First a random event is deleted. Next
the shortened story uses a "reader model" to determine whether the story is
still comprehensible. Some research uses a planner in the place of the
reader model; if the planner can fill in the missing elements of the story
then it is considered comprehensible.
1.a. Explain why this is a reasonable approach to the "reader model" or why
this is not a reasonable approach.
1.b. Describe what changes would need to be made to the planner (or planning
algorithm) to make it more acceptable as a substitute for a reader model.
1.c. Why might it not be a good idea to use a guess-and-check method?
KRR #2
After a hiatus of a decade or so, explanation appears to be returning as a
major topic of research in AI. This resurgence of research on explanation
partly due to the increasing ubiquity of computing in our society: computing
devices are everywhere but human typically do not understand the computations
and sometimes do not trust the devices. Another reason for the renewed focus
on explanation is the growing scale of computational software: some software
systems have become so large that apparently no one understands them or can
maintain them.
1. What are some of the differences in the requirements of explanation in the
above two situations?
2. Pick one of the above two situations. Give an example of an AI system that
generates (at least partial) explanations for that situation? Analyze the
knowledge and reasoning used by the system to generate explanations?
3. Will the same knowledge and reasoning work for generating explanations in
the other situation? Why?
KRR #3
In case-based reasoning, the retrieval process requires a function to
determine the "distance" between the current problem state and stored cases.
A common approach is to use domain-specific distance functions. An
alternative approach is to use a domain-independent measure of distance. One
way to do measure domain-independent distance is to use a full-blown
analogical reasoner such as Structure-Mapping Engine (SME) or Connectionist
Analogy Builder (CAB).
A) Under what circumstances should one use or not use analogical reasoners
such as SME and CAB in a case-based reasoning system?
B) What are the pros and cons of using analogical reasoners such as SME and
CAB in a case-based reasoning system?
=================================================
If one of your two areas is Cognitive Science, answer 2 of the 3 questions
below:
CogSci #1
The following is the abstract of a paper to appear in Brain and Behavioral
Sciences.
Darwin's mistake: Explaining the discontinuity between human and nonhuman
minds
Derek C. Penn, Keith J. Holyoak and Daniel J. Povinelli
Over the last quarter-century, the dominant tendency in comparative cognitive
psychology has been to emphasize the similarities between human and nonhuman
minds and to downplay the differences one of degree and not of (Darwin
1871). In the present paper, we argue that Darwin was mistaken: the profound
biological continuity between human and nonhuman animals masks an equally
profound discontinuity between human and nonhuman minds. To wit, there is a
significant discontinuity in the degree to which human and nonhuman animals
are able to approximate the higher-order, systematic, relational capabilities
of a physical symbol system (Newell 1980). We show that this symbolicrelational discontinuity pervades nearly every domain of cognition and runs
much deeper than even the spectacular scaffolding provided by language or
culture alone can explain. We propose a representational-level specification
of where human and nonhuman abilities to approximate a PSS are similar and
where they differ. We conclude by suggesting that recent symbolicconnectionist models of cognition shed new light on the mechanisms that
underlie the gap between human and nonhuman minds.
Question: Based on your cognitive science readings, give two examples of
cognitive tasks that require "higher-order, systematic, relational
capabilities of a physical symbol system."
Build an argument that while humans can do these cognitive tasks, other
animals cannot.
Does this imply that despite the biological continuity between humans and
nonhumans, there is a cognitive discontinuity between them?
CogSci #2
Barsalou has developed a theory of mental representation he calls "perceptual
symbol systems." He situates his theory in the movement advancing an
embodied account of mental representation. How does this account differ from
the traditional account of representation? What are its advantages? Its
disadvantages? What do you see as the implications of "perceptual symbol
systems" theory for an account of cognitive functions such as analogical
problem solving or mental modeling?
CogSci #3
Briefly, visual analytics refers to the use of external visual
representations of complex data in support of human analytical reasoning. As
an example, consider the task domain of investigative analysis. Let us
suppose that an investigator is dealing with a huge amount of data about
specific entities and specific events in a homicide case. The entities and
events in the data are related in many ways, e.g., spatially, temporally,
causally, intentionally, etc.
The investigator's task is analyze the data, understand the relationships
among the events and the entities, and construct one or more stories that
would explain (subsets of) the data.
Given our current understanding of human cognitive structures and processes,
how may an interactive environment visually represent the above data to
facilitate the investigator's task? According to cognitive science, what are
the useful levels of aggregation and abstraction that may promote the
investigator's problem solving?
====================================================
If one of your two areas is Machine Learning, answer 2 of the 3 questions
below:
ML #1
Figure 4.15 on page 117 of the Russell and Norvig book (2nd edition) shows an
example of a genetic algorithm (GA). One of your professors would argue that
a genetic algorithm is just a less efficient way of a local stochastic search
or simulated annealing. Relate each step (a-e) in the diagram to simulated
annealing. Why is step e (mutation) necessary? What part of the simulated
annealing algorithm might make it more effective or efficient than the
genetic algorithm described?
ML #2
In creating a handwriting recognizer for a computerized pen tablet, I assign
a fully-connected, ten state left to right HMM topology to each of the 52
cursive letters (capital and lowercase). Why is this topology unwise for
training? Describe two methods that I might use to determine a more
appropriate topology tailored for each letter.
ML #3
Reinforcement learning has become an increasingly popular framework for
control. Unfortunately, RL algorithms often do not scale particularly well.
There are several ways in which this problem is manifest, most obviously in
the case when there are many states, or when actions are at a very low level
of abstraction.
a) Typical approaches to addressing such problems in RL include function
approximation and problem decomposition. Compare and contrast these two
approaches. What problems of scale do these approaches address? What are
their strengths and weaknesses? Are they orthogonal approaches? Can they work
well together?
b) What are the differences between hierarchical and modular reinforcement
learning? Explain both the theoretical and practical limits of these
approaches.
=====================================
If one of your two areas is Perception, answer 2 of the 3 questions below:
Perception #1
This question relates to the following papers from the reading list. (1)
Brunelli, R. & T. Poggio (1993). (2) Viola and Jones (2001) and (3) Black &
Jepson (1996) and (4) Irani (1999).
(1) and (2) are really about image representations for object level matching
and (3) and (4) are about tracking. Are there any commonalities in the
approaches used in these efforts (between (2) and (3) is obvious, but please
explain it). Also, paper (1) talks about features vs. templates. How can
you use that to relate (3) and (4). In general for motion, what do you think
is better, features or templates. Where does motion analysis (optical flow)
show up in that? Why is optical flow fallen out of favor for tracking
recently? What methods are much wider use?
Perception #2
Optical Flow
Explain Michal Irani's use of subspace constraints in optical flow.
Specifically, what is really going on? Can you explain in laymen's terms why
there are such constraints? Then, how does she take advantage of them to
improve optical flow estimation? Finally, how come these constraints do not
hold for longer sequences, and can you think of subspace constraints that
would hold for longer sequences, perhaps subject to some assumptions...
Perception #3
Features for Categorization
Within the past ten years the computer vision community has made tremendous
progress in identifying effective features for problems in matching and
categorization. Give two examples of widely-used feature representations. One
of your example features should be suitable for the problem of finding
correspondences between images, as is needed for example in structure from
motion. The other example feature should be suitable for a categorization
problem, such as object or scene categorization. For each of your two feature
examples, explain briefly how it could be used to solve a specific perception
task. Most people would agree that the features in use today are
significantly more effective than those used 10 years ago. If you had to
identify the single most important insight that led to improved feature
representations, what would it be? Explain. Choose one of your two example
features and identify its most critical weakness (the property that, if it
could be improved, would make the biggest difference in improved
performance.) Explain your choice.
=========================================================
If one of your two areas is Planning & Search, answer 2 of the 3 questions
below:
P&S #1
You job is to create a hybrid planner for traveling to a conference. Going to
a conference involves getting tickets, driving to the airport, getting on the
plane etc. Each task may require distinct planning methods including
symbolic, geometric and probabilistic reasoning.
a) Choose three tasks involved in going to a conference: one for each type of
reasoning. Describe why that type of planning is most appropriate for the
task.
b) For each task in (a), present a simple representation of the domain
(pictures help). Make sure to include states, actions and a statement of the
start and goal.
c) For each task in (a), propose a specific planning algorithm that solves
the domain. Explain why that algorithm is the best choice. Focus particularly
on considerations related to completeness, optimality and efficiency.
P&S #2
Many modern motion planners focus on probabilistic completeness instead of
completenes. In these questions, consider both the theoretical and practical
challenges of probabilistically complete planning.
a) Give a specific example of a planning domain where probabilistic
completeness would be undesirable. What planning algorithm would you use in
this domain?
b) What kinds of domains would an RRT be most useful in? Give one example
where an RRT would perform better than a complete planner. Give one example
where an RRT would be preferable to a PRM. In each case, explain.
c) Suppose you have a probabilistically complete planner such as a bidirectional RRT. You wish to use it to plan from a start to a goal while also
going through an ordered sequence of via-points. In less than 10 lines, write
pseudo-code for an algorithm that generates the plan.
d) Is your algorithm in (c) complete? Is it probabilistically complete?
Explain why it satisfies these criteria or give a counterexample.
P&S #3
Motion Planning with Many Degrees of Freedom:
You are asked to create a motion planner for a robot with many degrees of
freedom such as a Humanoid robot or Mobile Manipulator that avoids obstacles.
a) What representation would you use for your state space?
b) What algorithm would you use?
c) Give at least two representations and compare them with regard to
completeness, optimality and efficiency.
d) Give at least two algorithms and compare them with regard to analogous
metrics.
e) Choose a representation and algorithm that you think would be most
effective. Suppose we add an additional constraint that your robot has to
maintain balance while it avoids obstacles. How would you incorporate this
constraint into your planner? Would it still have all the positive qualities
you described earlier?