Version Space Algebra and its Application to Programming by Demonstration Tessa Lau [email protected] Pedro Domingos [email protected] Daniel S. Weld [email protected] Dept. of Computer Science & Engineering, University of Washington, Box 352350, Seattle, WA 98195-2350 USA Abstract output of the learned function is not a discrete scalar, but a structured object (e.g., a sequence of actions, Machine learning research has been very a description of a scene, the text corresponding to a successful at producing powerful, broadly- speech stream, the 3D shape of a protein). As a re- applicable classification learners. However, sult, applying machine learning is largely a \black art," many practical learning problems do not fit with few formalized principles or domain-independent the classification framework well, and as a re- solutions available. In this paper we begin to address sult the initial phase of suitably formulating this problem by extending Mitchell's (1982) version the problem and incorporating the relevant space framework to more complex learning problems. domain knowledge can be very difficult and We propose the notion of a version space algebra in time-consuming. Here we propose a frame- which simple, restricted version spaces are combined work to systematize and speed this process, into more complex ones using operations like union based on the notion of version space alge- and join. This approach lets the application designer bra. We extend the notion of version spaces combine multiple strong biases to achieve a weaker one beyond concept learning, and propose that that is carefully tailored to the application, and thus carefully-tailored version spaces for complex to reduce statistical bias for the least increase in vari- applications can be built by composing sim- ance. We also begin to build a library of reusable ver- pler, restricted version spaces. We illustrate sion space components. We apply these components our approach with SMARTedit, a program- and the version space algebra in the construction of ming by demonstration application for repet- the SMARTedit system, a programming by demonstra- itive text-editing that uses version space alge- tion application for the text-editing domain. In many bra to guide a search over text-editing action cases, SMARTedit learns generalized macros that au- sequences. We demonstrate the system on a tomate repetitive text transformations after just a sin- suite of repetitive text-editing problems and gle training example. present experimental results showing its ef- fectiveness in learning from a small number Version space algorithms have not been widely ap- of examples. plied in practice, mainly because their extension to noisy data is non-trivial, and because the boundary- set representation they use is often not efficient enough 1. Introduction (although see the work of Norton and Hirsh (1992), Hirsh et al. (1997), and others). We expect that More often than not, the most difficult and time- similar problems will be felt in the broader range of consuming part of developing a machine learning ap- applications we extend the version space approach to plication is formulating the problem in terms amenable here. However, the introduction of version spaces was to currently-available learners (Langley & Simon, of enormous value in clarifying the concept learning 1995). Machine-learning research has largely focused problem, and in moving from a phase of building ad on classification, and although many sophisticated hoc systems with limited uses to the current generation classifiers have been produced, many (perhaps most) of powerful, widely-applicable classifiers. The goal of problems are not easily cast in this framework. For ex- this paper is to provide a first step in doing the same ample, in many domains (e.g., problem solving, vision, for a broader class of learning problems. Moreover, it speech recognition, computational biology) the desired may often be the case that some component version the properties of convexity and definiteness are neces- spaces in an application can be efficiently maintained sary and sufficient for a version space to be BSR. while others require heuristic search. Combining the Mitchell's approach is appropriate for concept learn- two should produce better results than approaching ing problems, where the goal is to predict whether an the whole problem in an ad hoc manner. example is a member of a concept. We extend the approach to any supervised learning problem (i.e., to 2. Version Space Algebra learning functions with any range) by allowing arbi- trary partial orders. We base this proposal on the In this section, we define our concept of a version space observation that the efficient representation of a ver- algebra: a method for composing together many sim- sion space by its boundaries only requires that some ple version spaces using algebraic operations such that partial order be defined on it, not necessarily one that the whole is also a version space. We first extend ver- corresponds to generality. The partial order to be used sion spaces to apply with any partial order (not just in a given version space is provided by the applica- generality, as in Mitchell (1982)). We then define the tion designer, or by the designer of a version space li- union, intersection, join, and transform operators over brary. The corresponding generalizations of the G and these extended version spaces, and describe the condi- S boundaries are the least upper bound and greatest tions under which the combined version space may be lower bound of the version space. As in Mitchell's ap- efficiently maintained. proach, the application designer provides an update A hypothesis is a function that takes as input an ele- function U(VS; d) that shrinks VS to hold only the ment of its domain and produces as output an element hypotheses consistent with example d. of its range. A hypothesis space is a set of functions We now introduce a version space algebra using these with the same domain and range. The bias deter- extended version spaces. We define an atomic version mines which subset of the universe of possible func- space to be a version space as described above, i.e., one tions is part of the hypothesis space; a stronger bias that is defined by a hypothesis space and a sequence of corresponds to a smaller hypothesis space. We say examples. We define a composite version space to be that a training example (i; o), for i 2 domain(h) and a composition of atomic or composite version spaces o 2 range(h), is consistent with a hypothesis h if and using one of the following operators. only if h(i) = o. A version space, VSH;D, consists of only those hypotheses in hypothesis space H that are Definition 1 (Version space union) Let H1 and consistent with the sequence D of examples. When a H2 be two hypothesis spaces such that the domain new example is observed, the version space must be (range) of functions in H1 equals the domain (range) updated to ensure that it remains consistent with the of those in H2. Let D be a sequence of training exam- new example. We will omit the subscript and refer to ples. The version space union, VSH1;D [ VSH2;D, is the version space as VS when the hypothesis space and equal to VSH1[H2;D. examples are clear from the context. In Mitchell's (1982) original version space approach, Hirsh proved that the union of two BSR version spaces the range of functions was required to be the Boolean is also BSR if and only if the union is convex and def- set f0; 1g, and hypotheses in a version space were par- inite. In contrast, we allow unions of version spaces such that the unions are not necessarily boundary- tially ordered by their generality. (A hypothesis h1 is set representable, by maintaining component version more general than another h2 iff the set of examples spaces separately; thus, we can efficiently represent for which h1(i) = 1 is a superset of the examples for more complex hypothesis spaces. which h2(i) = 1.) Mitchell showed that this partial or- der allows one to represent the version space solely in Proposition 1 (Efficiency of union) The time terms of its most-general and most-specific boundaries (space) complexity of maintaining the union is a lin- G and S (i.e., the set G of most general hypotheses ear sum of the time (space) complexity of maintaining in the version space and the set S of most specific hy- each component version space. potheses). The consistent hypotheses are those that lie between the boundaries (i.e., every hypothesis in the Definition 2 (Version space intersection) version space is more specific than some hypothesis in Let H1 and H2 be two hypothesis spaces such that G and more general than some hypothesis in S). We the domain (range) of functions in H1 equals the do- say that a version space is boundary-set representable main (range) of those in H2. Let D be a sequence (BSR) if and only if it can be represented solely by of training examples. The version space intersection, the S and G boundaries. Hirsh (1991) showed that VSH1;D \ VSH2;D, is equal to VSH1\H2;D. The considerations made above for the version space their respective data, it is not always the case that union also apply to the version space intersection. hA; Xi is consistent with the joint data. Although we have not yet formalized the conditions under which In order to introduce the next operator, let C(h; D) be joins may be treated as independent, Section 4 gives a consistency predicate that is true when hypothesis h several examples of independent joins in the PBD do- is consistent with the data D, and false otherwise.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-