Citation Context

...xt of PBD. Of course, missingvalues have been extensively studied in the broader statistics and machine learning fields. For example, [25] is a classic reference on incomplete data in statistics, and =-=[30]-=- explains how C4.5 deals with missing attribute values. Figure 12. A traditional wizard interface which pops up as a separate window to prompt for necessary parameter values. This form of mixedinitiat...

Citation Context

...rature, and no one has tackled missing-value problems in the context of PBD. Of course, missingvalues have been extensively studied in the broader statistics and machine learning fields. For example, =-=[25]-=- is a classic reference on incomplete data in statistics, and [30] explains how C4.5 deals with missing attribute values. Figure 12. A traditional wizard interface which pops up as a separate window t...

Citation Context

...ing example (i, o), for i ∈ Ih and o ∈ Oh, if and only if h(i) = o. A version space, VSH,D, consists of only those hypotheses in hypothesis space H that are consistent with the sequence D of examples =-=[26]-=-. When a new example is observed, the version space must be updated to ensure that it remains consistent with the new example. For example, we may define a ConstInt version space containing functions ...

Citation Context

...on technique; previous work has resulted in both successful applications, a general machine-learning framework for the problem, and an understanding of the expressiveness / sample-complexity tradeoff =-=[7, 21, 22]-=-. However, several problems remain with most PBD systems: • Considerable Domain Engineering: Substantial domain engineering is required to modify PBD systems to work with a new application. This need ...

Citation Context

...uring learning. The key question is how to define the consistency of hypotheses defined in terms of missing values. Borrowing a technique from model-based diagnosis, CHINLE uses constraint suspension =-=[8]-=-; all constraints referencing a missing value are automatically considered consistent. For example, a conditional hypothesis is considered consistent with a state transition, A → B, if the value of th...

Citation Context

...tomatically transforms these descriptions into version spaces, applying a novel, probabilistic bias. Specifying Interfaces We build upon the open-source SUPPLE model-based interfacegeneration toolkit =-=[11]-=-, which uses decision-theoretic optimization to render interfaces in a device-independent manner (Figure 1). For the purposes of this paper, however, the only pertinent aspect of SUPPLE is its functio...

Citation Context

...rom ConceptNet and TAP. User-Control of the Learning Process In order to control automatic generalization of procedures, users must understand the programmatic representations. Ko and Myers’s WHYLINE =-=[15]-=- answers users’ “Why?” and “Why not?” questions; they show that this capability greatly helps users, reducing debugging time. A recent study on supporting end-user debugging [14] shows that the greate...

Citation Context

...re 1). For the purposes of this paper, however, the only pertinent aspect of SUPPLE is its functional interface specification (FS) language, described below. Following earlier work on model-based UIs =-=[31, 13, 27]-=-, SUPPLE represents an interface functionally, e.g., specifying what capabilities the interface should expose, instead of how to present those features (SUPPLE’s optimization algorithm makes these ren...

Citation Context

...re 1). For the purposes of this paper, however, the only pertinent aspect of SUPPLE is its functional interface specification (FS) language, described below. Following earlier work on model-based UIs =-=[31, 13, 27]-=-, SUPPLE represents an interface functionally, e.g., specifying what capabilities the interface should expose, instead of how to present those features (SUPPLE’s optimization algorithm makes these ren...

Citation Context

...re 1). For the purposes of this paper, however, the only pertinent aspect of SUPPLE is its functional interface specification (FS) language, described below. Following earlier work on model-based UIs =-=[31, 13, 27]-=-, SUPPLE represents an interface functionally, e.g., specifying what capabilities the interface should expose, instead of how to present those features (SUPPLE’s optimization algorithm makes these ren...

Citation Context

...ly by constraints on the set of consistent hypotheses, such as the boundaries of the set relative to a partial order (one that is convex and definite [12], but not necessarily the generality ordering =-=[19]-=-). The next section explains how CHINLE automatically generates a VS, describes how conditionals are handled, and defines our probabilistic bias. Recursive-Descent Transformation Since the user explic...

Citation Context

... appropriate for the more sophisticated intended users of ECLIPSE; however, our approach allows CHINLE to continue to update the likelihood of all hypotheses, even after the user has intervened. PLOW =-=[1]-=- allows a different type of user input, using a spoken narrative to choose between possible generalizations of a demonstrated action. Some end-user programming work takes natural-language instruction ...

Citation Context

... Learning Approaches to PBD Many PBD systems are essentially rule-based expert systems, but some researchers have used a general machinelearning framework. Our work is based on Lau et al.’s framework =-=[17, 19]-=- and exploits their compact, factored representation of the version space. In Lau et al.’s SMARTEDIT system, the designer manually specified the VS algebraic description, but CHINLE generates it autom...

Citation Context

...hting the expected step 4 (set Paper Size) and filling in the anticipated value A4; see Figure 5. To indicate the system’s confidence in its predictions, CHINLE uses a 6-level sequential color scheme =-=[5]-=-, generated by COLORBREWER; its confidence in A4 is very high (72%), which maps to dark green. Recovering by Adding a Missing Step 4 Note that highlighting was similarly used to guide users in PBD sys...

Citation Context

... communicates the system’s confidence in its predictions. In addition to machine learning, researchers have suggested using common-sense knowledge bases to power PBD generalization. For example, Creo =-=[9]-=- is a PBD web browser, which allows users to create a general-purpose procedure from a single example; Creo’s generalization ability comes from ConceptNet and TAP. User-Control of the Learning Process...

Citation Context

...on technique; previous work has resulted in both successful applications, a general machine-learning framework for the problem, and an understanding of the expressiveness / sample-complexity tradeoff =-=[7, 21, 22]-=-. However, several problems remain with most PBD systems: • Considerable Domain Engineering: Substantial domain engineering is required to modify PBD systems to work with a new application. This need ...

Citation Context

...nstrated action. Some end-user programming work takes natural-language instruction even further. Tailor [4, 3] allows the user to modify a procedure with ordinary English commands. Sloppy programming =-=[24, 23]-=- lets users automate Internet browsing actions, again with English instructions. CONCLUSIONS This paper describes CHINLE, which (like FAMILIAR [29]) automatically generates PBD systems for application...

Citation Context

...is very high (72%), which maps to dark green. Recovering by Adding a Missing Step 4 Note that highlighting was similarly used to guide users in PBD systems such as Eager [7, Chapter 9] and DocWizards =-=[2]-=-.Figure 7. If the user desires, they can inspect the state of the learning process. Additional columns show the training data for each demonstration, and by expanding a step (e.g. Step 2) we see an o...

Citation Context

...nstrated action. Some end-user programming work takes natural-language instruction even further. Tailor [4, 3] allows the user to modify a procedure with ordinary English commands. Sloppy programming =-=[24, 23]-=- lets users automate Internet browsing actions, again with English instructions. CONCLUSIONS This paper describes CHINLE, which (like FAMILIAR [29]) automatically generates PBD systems for application...

Citation Context

...on technique; previous work has resulted in both successful applications, a general machine-learning framework for the problem, and an understanding of the expressiveness / sample-complexity tradeoff =-=[7, 21, 22]-=-. However, several problems remain with most PBD systems: • Considerable Domain Engineering: Substantial domain engineering is required to modify PBD systems to work with a new application. This need ...

Citation Context

...ace algebra, CHINLE handles inconsistent training data differently from most machine-learning systems, which strive to be tolerant of noisy data. Systems based on decision trees [29], HMM derivatives =-=[16]-=-, support-vector machines, or similar methods continue to predict the highest probability action, even when the user’s actions contradict all representable hypotheses. While noise-tolerant methods are...

Citation Context

...cases, version spaces may be represented efficiently by constraints on the set of consistent hypotheses, such as the boundaries of the set relative to a partial order (one that is convex and definite =-=[12]-=-, but not necessarily the generality ordering [19]). The next section explains how CHINLE automatically generates a VS, describes how conditionals are handled, and defines our probabilistic bias. Recu...

Citation Context

...s. In particular, one possible representation for String is a conditional, which is represented as a join of a version space for the condition being tested, and values for the true and false branches =-=[18]-=-. Lau’s framework allows the application designer to specify a preference bias by defining a probability distribution over the hypotheses in the hypothesis space. CHINLE uses a particularly useful dis...

Citation Context

...rted mixedinitiative PBD procedures. If TELS makes a mistake, the user is invited to enter a debugging phase which reverts to learning mode. Active-learning approaches, such as that of Wolfman et al. =-=[32]-=-, use the system’s understanding of its uncertainty to proactively determine which question might best be asked of the user in order to speed learning. We wish to incorporate these ideas into CHINLE. ...

Citation Context

... limitation of CHINLE is its requirement that the user manually segment the execution trace into demonstrated segments of fixed length. Other methods, such as IOHMMs [16], Augmentation-Based Learning =-=[28]-=- and Distributed Augmentation-Based Learning [6] relax this assumption, solving the alignment and generalization problem. To the best of our knowledge, partial learning has not been discussed in the P...

Citation Context

...ning algorithms, and user control of the learning process. Automatic Construction of PBD Systems Little work has been done on the automatic generation of PBD systems, but Frank and Foley’s early work =-=[10]-=- is an important exception. Unfortunately, the PBD systems it was able to generate had limited generalization capabilities, focusing only on the size and placement of 2D objects. Piernot and Yvon’s AI...

Citation Context

...o and Myers’s WHYLINE [15] answers users’ “Why?” and “Why not?” questions; they show that this capability greatly helps users, reducing debugging time. A recent study on supporting end-user debugging =-=[14]-=- shows that the greatest need for explanation falls in the Oracle/Specification category: figuring out whether a value was right or wrong and how to fix values. The DOCWIZARDS [2] system addresses the...

Citation Context

...ype of user input, using a spoken narrative to choose between possible generalizations of a demonstrated action. Some end-user programming work takes natural-language instruction even further. Tailor =-=[4, 3]-=- allows the user to modify a procedure with ordinary English commands. Sloppy programming [24, 23] lets users automate Internet browsing actions, again with English instructions. CONCLUSIONS This pape...

Citation Context

... user manually segment the execution trace into demonstrated segments of fixed length. Other methods, such as IOHMMs [16], Augmentation-Based Learning [28] and Distributed Augmentation-Based Learning =-=[6]-=- relax this assumption, solving the alignment and generalization problem. To the best of our knowledge, partial learning has not been discussed in the PBD literature, and no one has tackled missing-va...

Citation Context

...use it uses version-space algebra, CHINLE handles inconsistent training data differently from most machine-learning systems, which strive to be tolerant of noisy data. Systems based on decision trees =-=[29]-=-, HMM derivatives [16], support-vector machines, or similar methods continue to predict the highest probability action, even when the user’s actions contradict all representable hypotheses. While nois...