Dept. Colloquiua

CS 470   Artificial Intelligence,  Instructor:  Jeffrey Horn

COURSE ANNOUNCEMENTS (Wednesday, April 27, 2016)

What's REALLY New:

• Final Project added.  See below.

Not so New (Older Announcements):

• Quiz 2 will be in class this Thursday.  See below.
• Quiz One will be in class this Thursday.  See below.
• Welcome to class!
• This is where old announcements will go, so that you won't miss any news!

• Syllabus

• Lecture Notes

• Assignments

• Programming
• Tests and Quizzes

• Assignment 1:   Introduction to State Space Search (FWGC and Uninformed Search)
  • Assigned:   Thursday, January 14, 2016
  • Due:  TBA
  • Code for A1 is here (cabbage.scm has the Scheme code and plt-301-bin-i386-win32.exe installs the older version of Dr. Scheme that has been confirmed to work with the source code in cabbage.scm. )
  • TASKS:
    1. Graph the entire State Space for the FWGC problem, as defined in class.
       • Each vertex represents a unique state.   Label each vertex with the state descriptor using Scheme syntax.
       • Use directed edges (i.e, single-head arrows as arcs) between states, labeling each edge to indicate the type of move (i.e., "F" for farmer only, "W" for farmer and wolf, "G" for farmer and goat, etc.)
       • Draw all states, including "illegal" states that violate one or more contraints of the problem, above.   Draw an "X" through all illegal states.
       • DO draw all edges going INTO illegal states.   DO NOT draw edges FROM illegal states.
       • Indicate the original problem by circling the initial and goal states and highlighting the path from initial to goal (i.e., "(w w w w)" to "(e e e e)" ).
    2. Answer the following questions:
       1. What is the longest plan that you can generate using the FWGC Scheme code provided in class?  Give me the length of the plan, the plan itself, and the problem (i.e., pair of initial and goal states).
       2. What can you say about the symmetry of the state space for FWGC?
       3. Give another problem (i.e., pair of initial and goal states) in the FWGC state space that yields a plan of at least length four.   (Give me the problem and a plan that solves it!)
• • Assignment 2:   Introduction to Heuristic Search (Blocks World and Informed Search)
    • Assigned:   Tuesday, Jan. 26, 2016
    • Due:  TBA
    • Submission:  Hardcopy or email is fine.    Just write it up CLEARLY!   And use the notation introduced below.
    • Code for A2 (emailed) (block.scm. )
    • TASKS:   ( For all of the tasks below, assume the single goal state G = " ( (A) (B C) (D) ) "   )
      • Design your own heuristic for OUR Blocks World (the one in Scheme; see link above)
        • Your heuristic should be a function from blocks world states to a scalar value (actually, non-negative integers).  
        • Use the notation "h*(s)", where s is a state of the blocks world.  
        • The cost of a path or plan should be measured as the number of move operations (function calls) in the plan.
        • Your h*(s) should be an estimate of the cost of minimal cost path (plan) to get from s to the goal state G.
        • Your h*(s) should be admissible, which means it should be an underestimate of the actual cost
        • Your h*(s) should be as informed as you can make it!  (Well, that is your challenge.)
        • Be as specific and precise as possible in describing your function.   Feel free to use natural language, mathematical notation, and computer code (pseudocode or real code) to make clear what your function computes.   I will try to implement some of the heuristics (in Scheme!) for class on Wednesday (or perhaps the following Monday) so that we can test and compare each other's functions!  
      • Evaluate your heuristic on the following states (i.e., give me  h*(s₁ ),  h*(s₂ ),  and h*(s₃ )   ):
        1.   s₁ = ( ( A B C D) ( ) ( ) )
        2.   s₂ = ( (A) (B) (C D) )
        3.   s₃ = ( (B C) ( ) (D A) )
  • Assignment 3:   Neural Networks
    • Task 1:   Backprop with 2 hidden layer units
      • Here is that example I promised of the kind of image that I want from each of you.   This kind of learned concept is similar to the examples we were making in class.
      • BTW, if you look at the source code for the paint method you will see how the concept is displayed:  we simply paint each pixel in the displayed instance space by running that coordinate through the trained NN as an example to be classified.   If the NN outputs "1" then we color the pixel green, otherwise red.   Note that this is fast because our NN is tiny.   If we had trained a million-neuron network we might have to wait a long time to classify the entire instance space in the 300x300 pixel displayed area!   (Of course when we scale up the NN we would want to use faster software (i.e., NOT Java!) and/or hardware (like GPUs).)
    • Task 2:  Backprop with 3 hidden layer units
    • Task 3:
      • Train our BackProp network (here) on PARITY (odd or even).   You may choose the bit string length L to use.   You may only use half of the 2^L instances for training.   Send me your training set (two files).   Your grade will be proportionate to the score achieved on the test set that I USE on the BackProp network that I TRAIN on YOUR training files!
  • Assignment 4:  Genetic Algorithms for Optimization
    • Goal:  Become familiar with the GA and its application and behavior through optimization of a known, hard, and open problem (NYC Tunnels)
    • Tasks:
      1. Use the GA code in the folder "NYCtunnelsGA".  Run it on the NYC Tunnels problem with the following 10 random seeds:
         • 000xy, 111xy, 222xy, 333xy, 444xy, 555xy, 666xy, 777xy, 888xy, 999xy
         • where "xy" is your individual two digit number in the class (obtain from me).
      2. Experiment with the GA parameters until you get "good" performance using maximum 1,000,000 cost (that is, total number of fitness function evaluations, which you may upper bound as #gens * #popsize, unless you want to add the code to cache your function evaluations; that would give you more evolution!).  This means that you find the $38.79 million solution on at least one of your random seeds, and you find the second best known solution ($39.062 million on at least half of your random seeds).  
      3. Send me a text message or file with a list of the GA parameters that you used to complete Task 2 above, sufficient for me to REPRODUCE your results.
      4. Show me a plot of convergence for one of the runs.   You may print this or send me a screenshot, etc.   Label your plot (i.e., parameters, including random seed).   I want to see plots of on-line WORST, BEST, and AVERAGE fitnesses over time.  See example.

FINAL PROJECT   click here

TESTS AND QUIZZES

Quizes

• Quiz 1:    State Space Graphs
  • Thursday, Feb. 4, 2016 in class.   Open notes (including copies of the practice Q1 below!)
  • Practice quiz here, solution here.
• Quiz 2:  Stimulus-Response AI
  • Thursday, Feb. 11, 2016 in class.  Open notes.
  • Sample Quiz 2 (from 2012 course) is here.