INTELLIGENT   SYSTEMS   QUALIFIER
Spring   2008
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   #1:
The   female   solitary   wasp,   Sphex,   lays   her   eggs   in   a   cricket   that   she   has   paralyzed   and   brought   to   her   burrow.
The   wasp   grubs   hatch   and   then   feed   on   this   cricket.
The   wasp's   routine   is   to   bring   the   paralyzed   cricket   to   the   burrow,   leave   it   on   the   threshold,   go   inside   to   see   that   all   is   well,   emerge,   then   drag   the   cricket   in.
If   the   cricket   is   moved   a   few   inches   away   while   the   wasp   is   inside   making   her   inspection,   the   wasp,   on   emerging   from   the   burrow,   will   bring   the   cricket   back   to   the   threshold,   but   not   inside,   and   then   will   repeat   the   procedure   to   see   that   everything   is   all   right.
Experimenters   have   found   that   the   wasp   will   repeat   this   exact   procedure   many,   many   times   without   any   deviation.
(Part   1)   Invent   percepts,   actions,   and   a   production   system   that   the   wasp   might   be   using   in   behaving   this   way.
(Part   2)   Now   propose   one   other   architecture/process   that   may   explain   the   wasp's   behavior.
Show   details.
(Part   3)   Which   of   the   two   architectures/processes   above   better   explains   the   wasp's   behavior?
Why?
In   general,   how   can   we   find   whether   one   architecture/process   is   beter   for   explaining   behavior?
Core   #2:

In   the   four ‐ queens   puzzle,   we   try   to   place   four   queens   on   a   4x4   chess   board   so   that   none   can   capture   any   other.
That   is,   only   one   queen   can   be   on   any   row,   column,   or   diagonal   of   the   array.
Suppose   we   try   to   solve   this   puzzle   using   the   following   problem   space.
The   start   node   is   labeled   by   an   empty   4x4   array;   the   successor   function   creates   a   new   4x4   array   containing   one   additional   legal   placement   of   a   queen   anywhere   in   the   array;   the   goal   predicate   is   satisfied   if   and   only   if   there   are   four   legally   positioned   queens   in   the   array.
This   puzzle   can   be   solved   using   more   than   one   method.
(Part   1)   Invent   an   admissible   heuristic   function   for   this   problem   based   on   the   number   of   queens   placements   remaining   to   achieve   the   goal.
Note   that   all   goal   nodes   are   precisely   four   steps   from   the   start   node.
Use   your   heuristic   function   in   an   A*   search   to   the   goal   node.
Draw   the   search   tree   consisting   of   all   4x4   arrays   produced   by   the   search   and   label   each   array   by   its   value   of   g   and   h   functions.
Note   that   symmetry   considerations   mean   we   have   to   generate   only   three   successors   from   the   start   node.
(Part   2)   Now   we   can   try   to   solve   the   same   four ‐ queens   puzzle   as   a   constraint   satisfaction   problem.
The   constraint   graph   for   this   puzzle   contains   four   nodes,   where   each   node   represents   a   specific   column   on   the   4x4   chess   board.
Each   node   has   exactly   one   variable   that   represents   the   row   number   and   can   take   any   integer   value   between   1   and   4.
The   links   between   the   nodes   represent   the   constraints   of   the   puzzle
(no   queen   can   capture   any   other).
Draw   the   constraint   graph,   showing   all   the   nodes,   links,   and   variables.
Solve   the   puzzle   using   any   method   of   constraint   satisfaction.
(Part   3)   Which   method,   A*   or   constraint   satisfaction,   would   you   recommend   for   this   problem.
Why?
In   general,   what   criteria   are   appropriate   in   deciding   what   method   to   use   for   a   given   problem?
Core   #3:
In   1984,   Doug   Lenat   initiated   an   ambitious   project   called   Cyc   to   build   a   comprehensive   repository   of   all   human   knowledge   (see   www.wikipedia.org/wiki/Cyc).
This   would,   if   successful,   enable   AI   systems   to   reason   and   learn   at   a   human   level.
Initially   conceived   as   a   5   year   project,   Cyc   is   now   over   20   years   old   and,   although   useful,   still   far   from   complete.

In   the   meantime,   another   ambitious   project   to   capture   all   human   knowledge   has   had   phenomenal   success:   Wikipedia   (see   www.wikipedia.org/wiki/Wikipedia).
Just   over   5   years   old,   it   is   now   a   leading   reference,   albeit   sometimes   controversial.
Of   course,   Wikipedia's   knowledge   resides   in   text   form   and   is   not   usable   by   AI   systems.
(Part   1)   Could   Wikipedia   be   used   to   build   Cyc?
If   so,   explain   how.
If   not,   explain   why   not.
(We're   looking   for   a   technical   answer,   not   a   philosophical   one.)
(Part   2)   Could   Cyc   be   used   to   improve   Wikipedia?
If   so,   explain   how.
If   not,   explain   why   not.
(We're   looking   for   a   technical   answer,   not   a   philosophical   one.)
====================================================================================
If   one   of   your   two   areas   is   Perception,   answer   2   of   the   3   questions   below:
Perception   #1:
A   classical   problem   in   early   vision   is   texture   analysis   and   segmentation.
Historically,   it   has   been   argued   that   textures   could   be   modeled   as   a   collection   of   repeating   patterns   or   elements   known   as   "textons".
(A   common   example   is   an   image   of   a   golf   ball,   where   the   dimples   on   the   surface   constitute   the   textons).
Briefly   describe   a   texton ‐ based   solution   for   each   of   the   classical   problems   of   texture ‐ segmentation   and   shape ‐ from ‐ texture.
In   the   related   problem   of   texture   synthesis,   the   goal   is   to   synthesize   an   output   image,   given   an   input   image   that   contains   examples   of   the   desired   texture.
The   example ‐ based   approach   to   texture   synthesis,   as   popularized   by   the   paper   of   Efros   and   Leung,   has   proven   to   be   very   successful   in   practice.
Compare   and   contrast   the   example ‐ based   and   texton ‐ based   approaches   to   texture   analysis   and   synthesis.
What   are   their   respective   strengths   and   weaknesses?
Choose   one   of   the   two   problems   of   texture ‐ segmentation   and   shape ‐ from ‐ texture   and   sketch   out   a   possible   example ‐ based   solution   to   your   chosen   problem.
Be   as   specific   as   possible   about   your   algorithm   and   its   inputs   and   outputs.
Perception   #2:

You   have   the   fortunate   opportunity   to   spend   9   months   studying   in   Italy.
Excited   by   the   recent   progress   in   categorical   object   recognition,   you   decide   to   build   a   system   for   categorizing   buildings   according   to   their   architectural   style.
Given   an   input   photograph   of   a   building,   your   system   will   output   a   category   label.
Fortunately   you   have   access   to   a   hand ‐ labeled   database   of   building   images.
Sketch   out   a   partial   solution   to   this   object   recognition   task.
What   representation   of   the   building   appearance   would   you   use?
How   would   you   design   the   feature   space?
What   classifier   design   would   you   use?
How   will   you   deal   with   issues   like   segmentation?
Describe   how   your   system   would   be   trained   and   how   it   would   classify   a   novel   input   image.
What   would   be   some   examples   of   easy   and   hard   building   categories   to   distinguish   and   why?
How   would   the   sources   of   variability   in   building   design   affect   the   performance   of   your   system?
[Note   that   there   are   a   number   of   dimensions   along   which   buildings   can   differ   based   upon   their   architecture.
Some   examples   include   the   overall   shape   of   the   geometric   forms   that   make   up   the   building
(such   as   the   roof   shape   or   the   proportions   of   the   building),   any   decorative   treatments   applied   to   the   exterior,   placement   of   doors   and   windows,   and   choice   of   building   materials.
It   is   not   necessary   that   you   have   detailed   knowledge   of   classical   European   architecture   to   answer   this   question,   please   feel   free   to   use   any   reasonable   building   categories   in   describing   your   solution.]
Perception   #3:
You   are   charged   with   the   task   of   automatically   recognizing   dance   movements   from   video.
You   have   been   given   two   data   sets,   one   consists   of   videos   of   Fred   Astaire   dances,   the   other   contains   videos   of   bee   dances.
You   decide   to   use   HMM's   to   tackle   this   problem.
For   each   of   the   two   dance   domains   (Fred
Astaire   and   bees),   answer   the   following   questions:
(Part   1)   Describe   the   inputs   and   outputs   of   your   recognition   program
(Part   2)   What   do   the   states   of   the   HMM   represent?
How   many   states   will   you   have?
(Part   3)   What   are   the   features?

(Part   4)   How   will   you   define   the   topology?
(Part   5)   How   much   training   data   will   you   need   for   good   recognition   results?
(Part   6)   After   training,   can   you   suggest   a   visualization   technique   for   understanding   what   the   HMM   representation   has   captured?
(Part   7)   How   well   do   you   think   this   approach   will   work   for   recognizing   dances,   and   why?
Are   there   other   representations   that   might   work   better?
Why   or   why   not?
(Part   8)   You   notice   that   the   SIGGRAPH   deadline   is   tomorrow,   and   consider   re ‐ using   your   trained   HMM   model   as   a   means   to   synthesize   new   dance   sequences   for   a   graphics   application   (you   can   choose   either   of   the   two   dance   domains).
Is   this   a   viable   approach?
Why   or   why   not?
====================================================================================
If   one   of   your   two   areas   is   Robotics,   answer   2   of   the   3   questions   below:
Robotics   #1:
Having   grown   tired   of   watching   countless   Robocup   soccer   matches,   you   decide   to   invent   a   new   form   of   team   robot   competition   which   is   both   more   challenging   and   more   likely   to   inspire   the   new   generation   of   robotics   students.
Consider,   as   an   alternative,   the   classical   game   of   Capture   the   Flag   (CTF).
Briefly,   CTF   involves   two   teams,   each   of   which   has   a   base   location   with   an   associated   flag.
The   goal   for   each   team   is   to   reach   their   opponents   base,   remove   their   flag,   and   take   it   to   their   own   base.
The   first   team   to   accomplish   this   goal   wins.
Describe   a   possible   design   for   a   CTF   team   robot   competition,   which   would   be   analogous   to   the   current   Small   Size   League   (SSL)   competition   in   Robocup.
(You   can   assume   the   availability   of   a   hardware   platform   that   is   similar   to   any   recent   teams   in   the   SSL.)   Describe   the   fundamental   problems   in   the   coordination   of   robot   teams   that   must   be   addressed   in   solving   robot   soccer   and   in   solving   CTF.
Compare   and   contrast   these   two   competitions.
Which   one   is   more   challenging   and   why?

Robotics   #2:
Arkin   proposed   a   hybrid   architecture   for   robot   control   in   which   low ‐ level   control   is   handled   by
"behaviors"   that   map   perception   directly   to   action,   and   higher ‐ level   control   that   selects   which   behaviors   to   activate.
Usually   the   higher ‐ level   controller   is   a   state   machine   that   switches   between   behaviors.
Usually   there   is   a   "classical"   AI   planner   integrated   into   the   system.
(Part   1)   There   are   at   least   two   ways   a   planner   might   be   integrated:   1)   The   planner   acts   as   one   of   several   behaviors   to   suggest   directions   the   robot   should   go,   2)   The   planner   selects   groups   of   behaviors   to   turn   on   or   off.
(Part   2)   Both   of   these   methods   are   subject   to   failure.
Explain,   for   each   type   of   integration   how   the   system   might   fail.
Is   there   a   fundamental   lesson   to   be   learned   with   regard   to   these   failure   modes?
Robotics   #3:
Imagine   we   want   to   make   a   humanoid   robot.
We   do   the   best   we   can   to   give   the   robot   the   same   sensors
(eyes,   ears,   nose,   etc.)   and   affectors   as   the   human   body.
Next,   we   create   an   exoskeleton   that   fits   around   a   researcher   and   records   all   his   joint   movements   and   sensations   (sight,   sound,   etc.)   for   a   year.
(Part   1)   We   try   to   use   this   data   to   train   our   humanoid   robot   how   to   move.
What   are   the   pitfalls   of   this   approach?
What   are   the   benefits?
What   method(s)   would   you   use   to   train   the   system?
(Part   2)   We   try   to   use   this   data   to   teach   the   robot   how   to   respond   in   social   situations.
What   are   the   pitfalls   of   this   approach?
What   are   the   benefits?
What   method(s)   would   you   use   to   train   the   system?
====================================================================================
If   one   of   your   two   areas   is   Machine   Learning,   answer   2   of   the   3   questions   below:

ML#1:
Give   at   least   two   examples   of   unsupervised   dimension   reduction   methods.
Give   at   least   two   examples   of   supervised   dimension   reduction   methods.
If   the   end   goal   of   machine   learning   is   to   minimize   regression   error   or   classification   error,   should   an   unsupervised   dimension   reduction   method   ever   be   performed   as   a   pre ‐ processing   step   before   applying   the   appropriate   supervised   learning   method?
Why   or   why   not?
ML#2:
Discuss   the   relative   merits   of   loopy   belief   propagation   vs.
Gibbs   sampling   in   a   Markov   Random   field,   using   an   example   of   a   real   application   to   aid   the   explanation.
ML#3:
Researchers   have   generalized   Q   learning   to   multiagent   environments   by   replacing   the   Q   value   in   each   state   with   a   Q   table   mapping   joint   actions   to   expected   future   return.
In   this   setting,   the   "max"   in   the   Q   learning   update   is   replaced   by   a   game   theoretic   operation   (minimax,   Nash,   correlated   equilibrium).
Zinkevich   has   shown   that   there   exists   games   where,   in   each   state,   only   one   player   has   a   choice   and   yet   players   must   choose   actions   stochastically   to   achieve   the   unique   equilibrium.
Explain   why   this   fact   proves   that   the   multiagent   generalizations   of   Q   learning   are   not   guaranteed   to   converge   to   equilibria.
Does   this   sort   of   result   have   larger   implications   about   the   feasibility   of   methods   that   have   been   developed   for   multi ‐ agent   reinforcement   learning?
====================================================================================
If   one   of   your   two   areas   is   Planning   and   Search,   answer   2   of   the   3   questions   below:
Planning   &   Search   #1:

Some   have   made   the   argument   that   reinforcement   learning   is   simply   another   formulation   of   planning.
Would   you   agree   with   this   statement?
Why   or   why   not?
What   are   the   theoretical   and   practical   differences   between   classical   planning   and   the   approaches   used   to   solve   it   and   reinforcement   learning   and   the   approaches   used   for   solving   it?
Planning   &   Search   #2:
You   are   designing   a   planner   for   a   mobile   robot   in   a   mostly   static   environment   with   some   dynamic   obstacles.
The   planner   represents   the   world   as   a   binary   grid   of   obstacles   and   free   space   and   searches   it   with   A*.
Assuming   that   you   don't   know   how   the   dynamic   obstacles   will   move   design   the   best   possible
A*   heuristic   and   describe   how   you   would   compute   it.
How   would   you   improve   this   heuristic   if   the   dynamic   obstacles   moved   in   a   predictable   way?
Planning   &   Search   #3:
Under   what   circumstances   will   GraphPlan   stop   growing   the   graph?
In   STRIPS   planning,   prove   that
GraphPlan   is   complete:   it   returns   a   solution   or   terminates   in   finite   time.