CAPÍTULO IV. DESARROLLO DEL PROGRAMADE TRABAJO
4. DESARROLLO DEL MÉTODO IDEAL
4.1 PROPUESTAS PARA EL MEDIO AMBIENTE
Identifying Action-Sequence Shortcut Candidates
Two broad approaches can be taken to gather data on a user’s actions. The first requires a user to specify the actions he or she performs. Many
programming-by-demonstration approaches require the user to specify their actions, and examples of such approaches are given in Section 2.2.2 on page 21. The second approach continuously and passively records which actions a user makes, inferring the user’s intentions. Note that these two approaches are similar in nature to theintended and keyhole notions of plan recognition described by Cohen et al. [22]. Although the second approach has the benefit of not requiring direct user involvement it does introduce several difficulties.
The first difficulty is inferring the user’s intention from their actions. If actions are observed via low-level events such as key presses, then without knowing the context in which these actions are made, it may be ambiguous as to what high-level actions a user intends. The second difficulty is determining which action sequences made by a user should become an action-sequence shortcut. In the area of plan recognition, a similar problem of identifying pos- sible plans is often solved by determining the probabilities that certain action sequences will be observed and then selecting candidate plans based on these probabilities. An example of action-sequence probabilities for plan recog- nition in a user interface is presented in Waern [108]. Further approaches that use the probability of action sequences are presented in Section 2.2.2 on page 18.
Related to the issue of identifying candidate action sequences is the is- sue of determining which action sequences would make for “good” candidate action-sequence shortcuts. For instance, candidate action-sequences identi- fied for an application that is rarely used would be of little value to the user. Furthermore, some action sequences could represent tasks that the user would never want automated. An example might be a deletion task, where the user would always want to manually confirm a deletion. One means of determin- ing “good” candidate action sequences is by applying Blackwell’s attention investment model [13], which bases programming decisions on the cost, ben- efit and risk of “attentional” units to the user. Applying this model could provide a means of evaluating possible candidate action-sequence shortcuts.
5.2. PREDICTING USER ACTION SEQUENCES 131
Representing Action-Sequence Shortcuts
If a user manually defines an action-sequence shortcut, it can be expected that the user comprehends what the action-sequence shortcut will accom- plish. However, the aim of action-sequence shortcut induction is to remove the need for the user to specify an action-sequence shortcut and instead in- duce it from observation. As such, it is not possible to assume that a user will necessarily understand what an induced action-sequence shortcut will accomplish. Therefore conveying the meaning of an induced action-sequence shortcut becomes important.
Since an action-sequence shortcut automates a sequence of actions, a rel- atively simple representation scheme is to list a shortcut’s actions. However, such a scheme suffers from two problems. Firstly, this representation scheme is not compact. Secondly, determining how to represent actions may be dif- ficult. If actions are represented at the lowest level, such as key presses, then a user has to infer what the key presses will accomplish. Alternatively, if actions are represented at a higher level, such as tasks, then it requires the induction process to have some means of inferring such high-level tasks from key presses. The aim of many shortcut representation schemes is to employ a programming language that is easy for the user to comprehend. In particular, textual and visual programming languages have been studied as representational schemes. Burnett [15] discusses the strategies and issues of visual programming, as well providing examples of several techniques. Tex- tual representation schemes have also been employed, spanning the variety of programming languages available. A discussion of the merits and draw- backs between textual and graphical programming languages has been widely explored [87, 85, 106].
An alternative representational scheme is to represent a shortcut by the state it intends to bring about. Such representation schemes are often used in the planning domain, in which a plan is represented by its goal state. However, with this representation scheme there is still the difficulty of de- termining how to represent the goal state of a shortcut. Furthermore, by describing only the goal state, information regarding possible side-effects of achieving the goal state may not be conveyed to a user.
Predicting Action-Sequence Shortcuts
Apart from identifying action-sequence shortcut candidates and representing them, a further issue involves the presentation of action-sequence shortcuts to a user. Supposing there are many action-sequence shortcut candidates, it may not be feasible to visually present all action-sequence shortcut can- didates to a user. Furthermore, only one action-sequence shortcut will be appropriate to a user at any one time. The action-sequence shortcut induc- tion approach investigated in this chapter employs a learner to predict which action-sequence shortcut candidate to present to a user. In this regard, the action-sequence shortcut induction process can be seen as an approach for predicting which action sequence a user will perform. Similarities exist in the plan recognition domain, in particular, the plan matching stage. In tra- ditional plan recognition, the plan matching phase attempts to identify and explain a user’s current actions by matching their actions with a set of known plans. Various approaches have been employed to reason about which plans match a user’s intentions, including deductive approaches [56], probabilistic approaches [18] or a combination of both [91].
As with menu prediction, there exist several issues with employing a learner in this setting, including the computational burden that a learning approach would introduce on a mobile phone, the limited training examples provided by a user, the comprehensibility of the decisions produced by the learner, and concept drift. These matters are described in more detail in Section 3.2.2 on page 32.
Evaluating Action-Sequence Shortcut Prediction
As discussed in Section 3.2.1 on page 30, both efficiency and predictabil- ity are important characteristics of any adaptive user interface approach. These characteristics also apply to action-sequence shortcut induction. For instance, even though an action-sequence shortcut induction approach may increase the efficiency of an interface, if the action-sequence shortcut pre- sented alters frequently then the unpredictable nature of the approach rep- resents a cost to the user.
5.3. THE LEARNING SETTING 133