23.02.2015 Views

Machine Learning - DISCo

Machine Learning - DISCo

Machine Learning - DISCo

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

50 MACHINE LEARNING<br />

2.8. In this chapter, we commented that given an unbiased hypothesis space (the power<br />

set of the instances), the learner would find that each unobserved instance would<br />

match exactly half the current members of the version space, regardless of which<br />

training examples had been observed. Prove this. In particular, prove that for any<br />

instance space X, any set of training examples D, and any instance x E X not present<br />

in D, that if H is the power set of X, then exactly half the hypotheses in VSH,D will<br />

classify x as positive and half will classify it as negative.<br />

2.9. Consider a learning problem where each instance is described by a conjunction of<br />

n boolean attributes a1 . . .a,. Thus, a typical instance would be<br />

(al = T) A (az = F) A . . . A (a, = T)<br />

Now consider a hypothesis space H in which each hypothesis is a disjunction of<br />

constraints over these attributes. For example, a typical hypothesis would be<br />

Propose an algorithm that accepts a sequence of training examples and outputs<br />

a consistent hypothesis if one exists. Your algorithm should run in time that is<br />

polynomial in n and in the number of training examples.<br />

2.10. Implement the FIND-S algorithm. First verify that it successfully produces the trace in<br />

Section 2.4 for the Enjoysport example. Now use this program to study the number<br />

of random training examples required to exactly learn the target concept. Implement<br />

a training example generator that generates random instances, then classifies them<br />

according to the target concept:<br />

(Sunny, Warm, ?, ?, ?, ?)<br />

Consider training your FIND-S program on randomly generated examples and measuring<br />

the number of examples required before the program's hypothesis is identical<br />

to the target concept. Can you predict the average number of examples required?<br />

Run the experiment at least 20 times and report the mean number of examples required.<br />

How do you expect this number to vary with the number of "?" in the<br />

target concept? How would it vary with the number of attributes used to describe<br />

instances and hypotheses?<br />

REFERENCES<br />

Bruner, J. S., Goodnow, J. J., & Austin, G. A. (1957). A study of thinking. New York: John Wiey<br />

& Sons.<br />

Buchanan, B. G. (1974). Scientific theory formation by computer. In J. C. Simon (Ed.), Computer<br />

Oriented <strong>Learning</strong> Processes. Leyden: Noordhoff.<br />

Gunter, C. A., Ngair, T., Panangaden, P., & Subramanian, D. (1991). The common order-theoretic<br />

structure of version spaces and ATMS's. Proceedings of the National Conference on Artijicial<br />

Intelligence (pp. 500-505). Anaheim.<br />

Haussler, D. (1988). Quantifying inductive bias: A1 learning algorithms and Valiant's learning framework.<br />

Artijicial Intelligence, 36, 177-221.<br />

Hayes-Roth, F. (1974). Schematic classification problems and their solution. Pattern Recognition, 6,<br />

105-113.<br />

Hirsh, H. (1990). Incremental version space merging: A general framework for concept learning.<br />

Boston: Kluwer.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!