This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: Homework-6 Solutions Question 1 As discussed in class the concept of a medium built man can be expressed as an axis parallel rectangle. Therefore, it can be defined in terms of 4 real numbers: a, b, c, d such that: a ≤ height ≤ b, c ≤ weight ≤ d We assume that a consistent algorithm is available. Part I In this part we take as hypotheses 4 floating point numbers, each represented in terms of 4 bytes (32 bits). a How many hypotheses are in the hypotheses class? Answer: The number of hypotheses in the hypotheses class is: r = 2 32 · 4 = 2 128 b How many randomly obtained examples are needed for PAC learning with accuracy parameter epsilon1 = 0 . 1 and confidence parameter δ = 0 . 05? Answer: At least 918 randomly obtained examples are needed. m ≥ 1 epsilon1 ln( r/δ ) = 10 ln(2 128 / . 05) = 917 . 2 c What is the confidence level (what is δ ) if it is known that 1000 randomly chosen examples were used, and the desired accuracy is level is epsilon1 = 0 . 1? Answer: δ ≥ r (1- epsilon1 ) m = 2 128 × . 9 1000 = 5 . 95 E- 8 Answer can also be calculated based on formula m ≥ 1 /epsilon1 ln( r/δ ), which gives us δ ≥ 1 . 266 E- 5. This value is not as sharp as the one above. d What is the accuracy level (what is epsilon1 ) if it is known that 1000 randomly chosen examples were used, and the desired confidence level is δ = 0 . 05? Answer: δ ≥ r (1- epsilon1 ) m ⇒ (1- epsilon1 ) m ≤ δ/r ⇒ m ln(1- epsilon1 ) ≤ (ln δ- ln r ) ⇒ 1- epsilon1 ≤ e (ln δ- ln r ) /m ⇒ epsilon1 ≥ 1- e (ln δ- ln r ) /m ⇒ epsilon1 ≥ 1- e (ln 0 . 05- ln 2 128 ) / 1000 = 0 . 0876 Answer can also be calculated based on formula m ≥ 1 /epsilon1 ln( r/δ ), which gives us epsilon1 ≥ . 0917. This value is not as sharp as the one above. Part II In this part we assume that we are able to represent arbitrarily accurate axis parallel rectangles....
View Full Document
This note was uploaded on 12/03/2009 for the course COMPUTER E 770 taught by Professor Dr.mohammednizari during the Spring '09 term at Jordan University of Science & Tech.
- Spring '09