mgreen@cs.toronto.edu (Marc Green) (06/20/91)
From the tone of the posts, it's clear that many people view adaptive interfaces as some sort of HCI silver bullet. Well, there as many arguments to be made against adaptive interfaces as for them. The major problem is that the system's response is not completely predictable. People like machines to be slaves and to know that the same thing will have everytime. Moreover, an adaptive interface is going to make a lot of mistakes. Any diagnostic system, including ones which have to diagnose user intentions, makes errors. And Imagine the poor user when he tries to do something new for the first time and the interface keeps classifying as it some previously learned intention. You've got all the problems of HCI and all the problems of AI rolled into one. I'd put my money on customizable systems or systems which give advice on how to use more advanced features to accomplish a task more easily. This still puts the user in control, and that's where he wants to be. Marc Green Trent University
mcgregor@hemlock.Atherton.COM (Scott McGregor) (06/21/91)
In article <91Jun19.155043edt.6300@neat.cs.toronto.edu>, mgreen@cs.toronto.edu (Marc Green) writes: > The major problem is that the system's response is not completely > predictable. People like machines to be slaves and to know that the > same thing will have everytime. My experience with developing adaptive interfaces is that the above statement is an oversimplification of a more subtle point. As a result, it throws out the baby with the bathwater. My empirical observations do not suggest that people want slavish predictability, rather they want the ability to form useful habits. Normally this implies a certain amount of predictability, and repeatability, but it is not necessarily a slavish form. For instance, consider a scrolling list of items. An alphabetical layout may be more predictable, but a layout based upon putting the last used items at the top may be more convenient--and in cases where the same object is worked on for a long time, this allows the formation of a habit (grab item at top of the list) that might not be developed with the alphabetic list. Now in the case of the alphabetic list, you don't have to know any previous state to know where to look--so it is more predictable, but less useful if presumption of former state can be taken for granted. There are many other common examples of this sort of adaption based upon past state: E.g. the Macintosh root window remembers your previously open folders, IBM SPF remembers previous used field entries, etc. The more slavishly unchanging typical unix login sequence: "login:, password:, cd $HOME, shell prompt" is actually frequently more disatisfying to many users than the remembered Macintosh style boot-up, and this has lead to various different desktop approaches (X.desktop, LookingGlass, HP VUE) which are now trying to provide this level of adapatability as well. Now this is only a very limited level of adaptability (often so limited that many people neglect it), but even it is sufficient to demonstrate how slavish application of such generalizations concerning predictability can interfere with usability improvements. This extends to many other more complex adaptive situations, such as the EAGER hypercard macro generator discussed in the most recent CHI proceedings, or the Prescient Agent X windows system described in the May '90 HP Professional. Thus, a better way to state the general truth that Marc was trying to convey is: "if an adaptive interface interferes with habit formation, then it will annoy people, and reduce usability." > Moreover, an adaptive interface is going to make a lot of mistakes. > Any diagnostic system, including ones which have to diagnose user > intentions, makes errors. This is obviously true. However, it does not necessarily follow that the cost of the error will matter to the user. For instance, consider an adaptive interface used in an application to allow the user to choose a different a file to work on. The first time the application is used, the pull down menu for new file selection just allows the user to type in a new file name. After that, the user can still have the ability to select files by typing in the name, but can also choose from recently used files automatically added to the list. An additional filename added to the menu that the user doesn't care about is in one sense an error, but it doesn't interfere with the users actions so it is not annoying (additional examples of this sort of error are discussed in the Prescient Agents paper. > And Imagine the poor user when he tries to do something new for the first > time and the interface keeps classifying as it some previously learned > intention. If it gets in the way, then it is a bad interface, adaptive or not. But that doesn't mean that all adaptive interfaces interfere in this way, and so not all adaptive interfaces are bad. They aren't all necessarily good either. This is not an area to go about designing by mechanically applying generalizations -- there is still considerable individual talent components to designing good interfaces. > I'd put my money on customizable systems or systems which give advice > on how to use more advanced features to accomplish a task more easily. > This still puts the user in control, and that's where he wants to be. Advice and user directed customization aren't the only ways to leave the user in control. Sometimes the user's reactions to advice or the possibility of customization is frustration that the computer doesn't just do it. (Or at least offer to do it!). Many thoughtful adaptive interfaces do leave the user in control. My paradigm for an adaptive interface in our prescient agent work was M*A*S*H's company clerk Radar O'Reilley whose ability to "precache" answers to unasked questions facilitated his colonel's productivity instead of interfering with it. Adaptive interfaces can also be more like a "let me do it 2 year old" too, but they don't have to be, and shouldn't be judged solely on the worst case situation alone. Scott McGregor Atherton Technology mcgregor@atherton.com
de5@ornl.gov (Dave Sill) (06/21/91)
This is my second attempt to reply to this article, so bear with me if you somehow saw the first. In article <91Jun19.155043edt.6300@neat.cs.toronto.edu>, mgreen@cs.toronto.edu (Marc Green) writes: > >The major problem [with adaptive interfaces] is that the system's >response is not completely predictable. People like machines to be >slaves and to know that the same thing will have everytime. Moreover, >an adaptive interface is going to make a lot of mistakes. Any >diagnostic system, including ones which have to diagnose user >intentions, makes errors. And Imagine the poor user when he tries to >do something new for the first time and the interface keeps >classifying as it some previously learned intention. You've got all >the problems of HCI and all the problems of AI rolled into one. I think you're talking about *predictive* interfaces, not *adaptive* interfaces. The way I see it, there are at least the following major different types of interface. STATIC: The command set doesn't change from user to user or session to session, although small changes may occur as the software is refined. CUSTOMIZABLE: The command set can be altered by the user to include new commands or modify the behavior of commands or the appearance of the interface. The IF will tend to vary widly from user to user but remain fairly constant from session to session once customization is accomplished. ADAPTIVE: The command set and appearance of the IF are adjusted automatically by the software in an attempt to streamline the IF based on a particular user's usage patterns. For example, frequently selected menu items might migrate toward the tops of the menus, or new compund commands might be created and advertized that implement frequently executed command strings. PREDICTIVE: The software attempts to predict the user's intended sequence of actions before it's fully specified. For example, if a user always types "ls" after "cd", the IF might automatically execute the "ls". >I'd put my money on customizable systems or systems which give advice >on how to use more advanced features to accomplish a task more easily. >This still puts the user in control, and that's where he wants to be. If the interface is implemented well, the user will be in control, regardless of whether it's adaptive or predictive. -- Dave Sill (de5@ornl.gov) Tug on anything in nature and you will find Martin Marietta Energy Systems it connected to everything else. Workstation Support --John Muir