CPSC 633-600 Read-Only Bulletin Board

Last modified: 2/9/06, 11:27AM (Thu)

This page will have all relevant email transactions with the students regarding the AI course, so that everyone has equal information regarding the course material.

Newest transactions will be posted on the top. Regularly view this page to see what's going on between other students and the instructor.

All sensitive material such as your name, email address, etc. will be removed, and also any part of code that is not relevant to the discussion will not be uploaded here.


Article List

Date: 4/18 Title: Miniproject teams/topics
Date: 4/08 Title: Miniproject info
Date: 4/08 Title: Quiz 3 topics
Date: 3/05 Title: Quiz 2 topics
Date: 2/09 Title: Quiz 1 topics
Date: 1/12 Title: Welcome
Date: OLD 4/07 Title: Mini project details
Date: OLD 4/04 Title: HW3 corrections and clarifications
Date: OLD 2/15 Title: HW1 plotting the error curve
Date: OLD 2/10 Title: Dealing with missing attribute values
Date: OLD 2/09 Title: HW1 clarification


Articles

Date: 4/18 Title: Miniproject teams/topics

Teams and topics

  • Yoon Bai and DonhHyeop Han: GA simulaiton with dominant/recessive relationship
  • Amanda Lampton: Morphing space vehicles
  • Sizhuo (Andrew) Shi and Yuan-Teng Cheng: Path exploration
  • Daniel Eng and Xiao-lin Wei: OCR
  • David Seaburg: Avoiding moving obstacles with GA
  • Hassan A. Kingravi: Individual detection in movies using combined dimensionality reduction and k-NN.
  • Aditya Ramgopal and Vijay Kumar: Naive Bayes for document classification.
  • Belita Gopal and John Hamilton: Learing term weights for information retrieval using BBN.
Date: 4/08 Title: Miniproject info

Code-base

  1. Backpropagation: backprop-1.6.tar.gz (C++ code -- unix)
  2. Neuroevolution: ga.m (Octave code)
  3. SIDA: sida-nat.tar.gz (Octave code)
Final miniproject details 1. Pick a code-base - YC's backpropagation code (in c++) - YC's neuroevolution code (in octave) - YC's SIDA code for sensorimotor semantics (in octave) - Your own code - Reinforcement learning code (from homework) - Other - A third-party open source 2. Formulate your research problem - Pick task - Locate data set (if needed): see, e.g., UCI ML repository - Design experiments 3. Proposal - Team members: max 2 per team. - What is the research problem? - Why is it important/interesting? - What are other people's approaches? - What are the limitations of those approaches? - What is your approach? - What experiments will you do? - What are the expected results? - Submit by 4/16, in class. 4. Presentation - 20-25 minutes. - Present final or interim results 5. Final report - 4-5 page, single space report. - Due by 5/4, 3pm, HRBB 322B.
Date: 4/08 Title: Quiz 3 topics

You may bring 2 sheets of notes to the quiz.

Quiz 2 topics:

1. Genetic algorithms
	- Basics, plus selection strategies
	- Importance of representation (e.g., GABIL)
	- Analysis using schemas
	- Baldwin effect

2. Evaluating hypotheses
	- Sampling distribution of the mean
	- Confidence interval
	- Difference in error of two hypotheses
	- Comparing learning algorithms

3. Bayesian learning
	- Bayes theorem
	- Application of Bayes rule in analysis of learning algorithms
	- MAP hypothesis
	- Least square and Max Likelihood
	- MDL and relation to MAP
	- Bayes optimal classifier (different with MAP)
	- Gibbs sampling
	- Naive Bayes
	- Bayesian belief network
		- Basics
		- Monte Carlo inference
		- Gradient ascent for ML learning of cond. prob.
	- EM algorithm

4. Computational learning theory
	- Sample complexity
		- epsilon-exhausting thoerem
		- PAC learning
		- bound on sample complexity
		- VC dimension and sample complexity
	- Mistake bound
		- Find-S
		- Halving algorithm

Date: 3/05 Title: Quiz 2 topics
Quiz 2 topics:

1. Decision tree learning

   - relationship between entropy, uncertainty, and surprise
   - relationship between information gain and transition from
     one state to another
   - decision tree learning algorithm
	- choosing the best attribute
	- inductive bias

2. Reinforcement learning

   - value function V(s)
   - optimal policy pi
   - optimal value function V*(s)
   - delta(s,a) and r(s,a)
   - difficulty of V(s)-based approach (no knowledge of delta and r)
	- how does Q-learning overcome that difficulty?
   - Q-learning
	- deterministic
	- nondeterministic
	- [NEW] how to select actions given the current estimate of Q? 
   - TD(lambda)
	- role of lambda
	- equivalences (when lambda = 0 or lambda = 1)
Date: 2/09 Title: Quiz 1 topics
Quiz 1 topics
  • Inductive learning, inductive bias
  • Choice of target function
  • Representation of a hypothesis
  • Hypothesis space
  • Inherent structure in the hypothesis space and how search can be conducted efficiently
  • Version space
  • Concept learning: Find-S, list and eliminate, candidate elimination, convergence conditions
  • Perceptron: geometric interpretation, limitations, learning rule, gradient descent
  • Backpropagation: why it overcomes perceptrons, geometric interpretation, gradient descent (hidden to output and input to hidden), recurrent networks, hidden layer representations, limitations
  • Overfitting in inductive learning
Date: 1/12 Title: Welcome




Welcome to CPSC 633 Machine Learning. The articles below are from
the last time I taught 633 (Spring 2006), and would generally be helpful so I 
kept it here. New articles will be appeded at the top.





Date: OLD 4/07 Title: Mini project details
Submit a hardcopy of your mini-project proposal by Monday 4/10, in class. The proposal should contain:
  • Names and emails of participants.
  • Title of the project
  • A short, one-paragraph description of the project.
Presentation and final report:
  • Presentation:
    • will be roughly 15 minutes for each team.
    • You may present preliminary results.
  • Final report:
    • 5 to 6 pages, in a conference paper format of your choice. Single-spaced.
    • Due 2 days after the final exam (final exam is 5/8, 10:30am-12:30pm).
Date: OLD 4/04 Title: HW3 corrections and clarifications
1. For problem 2, use the same grid and reward as problem 1.

2. For problem 2 in general, checking for convergence in the way it 
   is described may give mixed results. You may 
	
	a. set a fixed max number of iterations and run all three policies
	b. compare the resulting Q(s,a) and the analytical results.

3. For the probabilistic policy in problem 2 (item 3), the two k values
   should be 2 and 5, not 1 and 5.

4. For problem 3, the subscript "t" represents the t-th visit to 
   (s,a).
Date: OLD 2/15 Title: HW1 plotting the error curve
Producing the error plot seems to be an issue. You may omit the intermediate steps, and just report the final error on the training set and the test set.
Date: OLD 2/10 Title: Dealing with missing attribute values
> I am having a little problem in finishing up the homework.
> It's working fine with the 'PlayTennis' data, but not with the 'Breast
> Cancer' data. I found out this is because of the case when Attributes or
> Examples_vi is empty(line 8 and 16 of the ID3 pseudocode in the textbook
> page 56.)
> In those cases, it says to add a node with label='most common value of
> Target_attribute in Examples'. This is not + or -, but one of the values of
> the attributes, then how can it make a decision in this case? And what is
> the 'most common' value of the attributes?

        There can be several strategies. You can simply replace the
        missing attribute value with the most commonly occurring
        value for that particular attribute. This can be done in some kind
        of preprocessing stage of the data.


Date: OLD 2/09 Title: HW1 clarification
Hi everyone,

For problem 1, item 2, you don't need to provide a formal proof. Just state your
intuition.

As for the breast cancer database, you may experience problem if you copy and paste
the url from the pdf file.

Use this url in that case:

  http://www.ics.uci.edu/~mlearn/databases/breast-cancer-wisconsin/

Yoonsuck

$Id: board.php,v 1.4 2003/09/04 21:56:27 choe Exp choe $