Designing for Human-Agent Interaction

d e Interacting with a computer requires adopting some metaphor to guide our actions an xpectations. Most human-computer interfaces can be classified according to two r t dominant metaphors: agent or environment. Interactions based on an agent metapho reat the computer as an intermediary which responds to user requests. In the environd ment metaphor a model of the task domain is presented for the user to interact with irectly. The term "agent" has come to refer to automation of aspects of human com. N puter interaction such as anticipating commands or autonomously performing actions orman’s 1984 model of HCI is introduced as reference to organize and evaluate i research in human-agent interaction. A wide variety of heterogeneous research involv ng human-agent interaction is shown to reflect automation of one of the stages of r action or evaluation within Norman’s model. Improvements in HAI are expected to esult from more heterogeneous use of methods which target multiple stages simultaneously. Agent vs. Environment s a Interacting with a computer requires adopting some metaphor to guide our action nd expectations. Most human-computer interfaces can be classified according to two r t dominant metaphors: agent or environment. Interactions based on an agent metapho reat the computer as an intermediary which responds to user requests. In the environd ment metaphor a model of the task domain is presented for the user to interact with irectly. Command line and screen editors provide a simple contrast between these ) i approaches. To change a word in a command line editor such as "ex" you must: 1 nstruct the editor to move to the appropriate line number then 2) enter some command o v such as s/old/new (locate string "old" then substitute for string "old" string "new"). T iew this change in the context of the surrounding text would then require composing e c a command such as p .-10,.+20 (print from 10 lines before to twenty lines after th urrent location). This same task is far easier with a screen editor where you simply y b move the sprite using mouse or cursor key to the offending word, delete it perhaps b ackspacing over it, and type in the replacement. This keystroke superiority of screen I over commandline editing is a well known (Card, Moran, and Newell 1983) HC 2 This research was supported by ONR grant N-00014-96-1-122 result. If the task were changed to "change every occurrence of old_word to s " new_word" the relative advantage is reversed and an instruction to an agent such a g/old/s/old/new" (locate string "old" then substitute for string "old" string "new" for m g=every occurrence of string "old") is far simpler than scouring the surrogate docu ent for occurrences of "old_word", erasing each, and typing "new_word" in its place. i For this example the character of predictable errors will differ as well; the subject nteracting directly with the document is likely to miss some occurrences of c "old_word" while the subject issuing the global command may suffer unintended onsequences such as changing "not_old_word" to "not_new_word". l r In practice, the better features of line editors such as string searching and globa eplace have almost always been retained in screen oriented editors, leading to interp faces in which indirect requests can be issued to perform tasks for which direct mani ulation proves too cumbersome. The distinction between agent and environment metaf phors is not identical to that between agent-based and direct manipulation-based inter aces which has been much debated (Schneiderman and Maes 1997). The t agent/environment distinction reflects the semantics (action vs. request) of the interac ion rather than its syntax (command line vs. button press). The binocular icon search e b button found on Netscape browsers, for example, uses the affordances of a pushabl utton to advertise its availability and means for initiating search but leaves the task of l locating a string to an "agent" rather than requiring the user to search the text line by ine. Task actions communicatable using an environmental metaphor are a proper subi set of those which could be specified to an agent and are just those tasks such as conic desktops, text editing, draw programs, or geographical information systems r which can provide a clear, literal correspondences between task domain and onscreen epresentation. The power of this approach which provides advertisement and unique d identification and selectability of available objects and actions is reflected in the ascen ence of graphical user interfaces (GUI’s). The value of the agent metaphor to e a interaction only becomes apparent when objects are not present or fully visualizabl nd actions are repetitive, delayed, or poorly specified. The distinctions between agent a and environment based HCI are very similar to those between manual and automated ction in the physical world. It is much simpler for us to drive a car or set a table a m than to instruct a robot to do so, yet we would rather adjust a thermostat or program illing machine than repeatedly performing these actions by hand. While the coma puter offers the ultimate in flexible automation, instructing it do what we wish may be rbitrarily hard for humans as demonstrated by the difficulty experienced in using tradi itional programming and scripting languages. The growing popularity of "agent-based" nteraction reflects the emergence of an increasingly powerful and complex computing r u environment bringing with it desires to perform flexible tasks involving multiple o nknown objects by users who do not wish or may not have the ability to program. e r Norman’s (1988) ecological model of HCI will be reviewed and used to organiz esearch in human-agent interaction. The premise is that software agents are intended s b to automate repetitive, poorly specified, or complex processes by bridging the gulf etween a user’s desires and actions which could satisfy them. Tasks, domains, and r interaction methods will be categorized according to the uncertainties they bring to o reduce at stages in this model. A maturing paradigm of human agent interaction is s a envisioned in which adaptation, user profiles, demonstration, and scripting are used a ppropriate to facilitate human-agent interactions. l D Reference Mode on Norman (1986) characterizes human-computer interaction as the problem of i bridging twin gulfs of execution and evaluation. The execution side of the cycle nvolves translating a goal into a sequence of actions for achieving that goal. The e a evaluation side involves using feedback from the domain to compare the result of th ction to the goal. The model is cybernetic rather than logical in that it emphasizes c feedback and incremental action rather than problem space search or planning. A cru ial feature of this model is that tampering with either side of the loop can lead to y f detrimental or unanticipated results. If the execution side is automated the human ma ail to observe effects of actions and be unable to correct errors or modulate ongoing e e behavior. If the evaluation side is automated the human may be unable to track th ffect of actions and adjust to their results. Norman proposes seven stages of action in this model to link the user’s goals to s i the world. The stages of execution are: forming an intention to act, translating thi ntention into a planned sequence of actions and executing that sequence. The stages t o of evaluation are perceiving the state of the world, interpreting this perception in ligh f prior action and evaluating that change with respect to initial goal. The gulfs refer i to the interface/metaphor which separates the user’s goals from the application domain n which they must be realized. An example of this form of analysis is shown below: _ ________________________________________________________________________ Changing the format of letter (Norman 1986, p. 44) _ ________________________________________________________________________ Goal improve the appearance of the letter e Intention change the paragraph style from indented to blocked; replac all new-paragraph commands with skip-line commands E Action Specification sub/.pp/.sp/whole xecution type "sub/.pp/.sp/whole" n I Perception each line of text begins at left margi nterpretation the paragraphs are blocked ? _Evaluation the letter now "looks better" ________________________________________________________________________ G Although approximate rather than precise and proposed in the context of early UI’s Norman’s model provides a useful reference for analyzing human computer a interactions of all forms because it identifies the cognitive processes and linkage mong them which must be supported for human-computer interactions to succeed. s h An agent is defined to be a program which automates some stage(s) of thi uman information processing cycle. This definition does not extend to software-only i agents found in multiagent systems and excludes human-computer interactions involv ng simple direct manipulation actions or explicit command line requests. e Figure 1 shows the Norman reference model and ways in which the involved cognitiv processing might be automated. Automated processes are indicated in italics within h a dashed boxes. Serial (non looping) automation strategies range from direct aiding suc s action sequencing or attentional filtering to executive functions such as anticipatory n m action or prioritized filtering. Agents which interact continuously with the domai ay be semiautonomous with or without feedback or may perform their assistive funci tions completely independent of the user. These categories could be elaborated to ncorporate the transparency of automated processes and other aspects likely to affect t human-agent interactions, but these nine should suffice to suggest some of the strucural complexities and concerns which should inform the design of agents to assist a human users. A rule of thumb would be that any serial configuration which automates complete side of the loop isolates the user and may lead to breakdown as may semiautonomous configurations which do not explicitly supply appropriate feedback. As shown in Figure 1 automation can occur through either semiautonomous r e processes set in motion by the user or through by-passing stages of execution o valuation which the user must otherwise perform. An agent which searches several . T databases for a favorable price would be an example of a semiautonomous process his form of closed loop automation is found in process industries where operators r a establish a variety of setpoints for valves, breakers, and proportional controllers. Fo utomation of this sort to succeed the user needs a relatively detailed mental model of ) a the domain and what the automation is to do in order to "program" it (transparency nd subsequently would benefit from good displays and methods for Goals Domain Intent Plan Interpret Execute Perceive Evaluate Goals