Mache Learni An Arti f dal [Vhel sk (983 gt Inte Usgene App poath Carbonell Mitchell. 15 LEARNING EFFICIENT CLASSIFICATION PROCEDURES AND THEIR APPLICATION TO CHESS END GAMES ABSTRACT nis deating with rae numbers of exam another developed resulls, ‘The paper ends ‘specification of relevant system is observed 15.1 INTRODU ON ‘ules discovered in this way as compare ‘ke same task, (in one case, the clas ‘est alternative method that I could devise.) is five times as fast as 463r “ CHAPTER 15: LEARNING EFFICIENT CLASSIFCATION PROCEDURES | QUINLAN a5 152 THE INDUCTIVE INFERENCE MACHINERY tee shown ‘dark’ and “rea sections, The first, Set jing System (C15). this model used to describe object ery of the attributes th466 Figure belonging in such eases the However, itis generally desi were not used in set of example: two object CHAPTER 15: LEARNING EFFICIENT CLASSIFCATION PROCEDURES r short, blond blue: + tal blond, Be: + tributes are inadequate for fe that the troe be able to classify objects wh ‘QUINLAN dark red nveyed by this message. and p~ tespectively, the expected informat p* logs pt — po 5, $0 that p+ becomes the propo MIC) to denote this eal fontent of a message from o deci 4) = 0. Now consider as before the po ‘st next. The paral decision tee is 9 Skribowe A are mnt mn set C of objects we can approximate these probabi 487 blond on of objects in C n of the expected a set C of objects, and define Mi ‘le choice of A as the ywn in Figure $5-4. The values Aj of268 CHAPTER 15: LEARNING EFFICIENT CLASSIFCATION PROCEDURES attribute A ea ea C1 C2 igure 154: til decision toe Toil the example given earlier M(C)= — 378 logy V8 ~ Si anch: (3 ~ 2/3 log 23 = 0.918 bits fon content = 5/8 0.971 + 3/8 * 0.918 this ateibute is: 0.954 — 0.951 = 0.003 bits which is ne previously. The branches the n. wh “objects und so requires 1 = 38 0+ UBM O + TT We have: en igible, The tree arising fromm testing the second ‘with 3 objects) and "red" he branch for “blond” co 954 - 0.5 = bute was giver object) requite ned 2 “plus” and 2 0.454 ‘QUINLAN 469 7~ tl, on, brown: = shot blow, + tie =| Short arte = short, blond, brown: = 1s on the set of objects C (such as determining the ihjects with valuc A; of altribute A) can be performed clticien practice that C has to be kept in fast memory. What happens ? One way around the difficulty is given by the version fh C is digested one object at two sets $ and G specific and far; these lecision tree random a subset of the given repeat © form a rule to ex © Find the exceptions t0 0 fo the n Uni there are no exceptions to the rule the current window is rule in the remaining instances the current window and the exceptions to new window le generated frot it ends when a nile is formed that has no excep J fall of C. Two diferent ways of forming 2 new window have been Wied Se ist, the current window is enlarged by the ad 1S. and 50 the window grows ‘als with a ication probh 2.000 objects for which a correct des 19794}. The main findings were: mm tree contained 48 nodes (Quinlan, La70 CHAPTER 15: LEAANING EFFICIENT CLASSIFCATION PROCEDURES | QUINLAN. S oa The converge rapidly; typically only 4 iterations were required to 1 eorreet decision tee. ‘#1 was possible to develop a correct tree from & only as ion of the 2,000 objects. ocess was nol very sensitive to parameters such as dow size. 474,000 ‘The firs tributes. 18 of these were window containing initial win the difficulty of the problem as defined by the ven in the introduction, have enabled 13 to discover correct jon problems, ‘These features, particularly the trees for some very large clas the only move the black ki 15.3 THE LOST N-PLY EXPERIMENTS, she rook IDS has been to discover classification rules for part of ik versus (black) king-knight. The relations com for n=2 and n=3; the 4-ply One application of the end game (white) ki last O-ply if and only if a the king is in checkmate, or b, the knight has been captured, the position is not stalemate, the whit rook has not been captured and the black king cannot retaliate by capturing it ibutes was meant task. For ex if the position was of one imtended to detect some ‘As expected, these new the classification problem than were rie predecestors. They had the effect of lost n= I ply. Tost n-ply (n even) iff all possible black moves ply n and S0-move rules of chess, but are quite IF attributes ean be found ion task and that are also relatively cheap to com are adequate f pute, then the «| bee faster than the minimay search of the gare ee. n_ways of placing the four pieces to form # this re the knight of whit ate). use, for instance, the white king ei fon.) These counts include many s)® metric variants of essentially the same position, and when these are removed tf | For the purposes Parison, all methods were implemented in Pascal