Information about http://www.ai.sri.com/pubs/files/1440.pdf
To appear in the Proceedings of the 2007 AAAI Spring Symposium on…
Tags: aaai spring symposium, artificial intelligence center, commitments, developing systems, initial thoughts, intelligent systems, karen myers, lems, menlo, menlo park, menlo park ca, nysmith, ond, position paper, proactivity, prob, research objectives, routine work, sri international, tance,Pages: 3
Language: english
Created: Mon Jan 29 09:04:43 2007
Display cached document
Page 1
Page 2
Page 3
To appear in the Proceedings of the 2007 AAAI Spring Symposium on Intentions in Intelligent Systems 1
Proactivity in an Intentionally Helpful Personal Assistive Agent
Karen Myers and Neil Yorke-Smith
Artificial Intelligence Center, SRI International, Menlo Park, CA 94025, USA
{myers,nysmith}@ai.sri.com
Abstract Our research objectives in this area are as follows. First,
we want to understand the types of proactive behavior that
The increased scope and complexity of tasks that people per- would be helpful to incorporate into an assistive agent. Sec-
form as part of their routine work has led to growing inter- ond, we want to characterize how an agent can best reflect
est in the development of intelligent personal assistive agents
over possible actions and current commitments (both user
that can aid a human in managing and performing tasks. One
desired capability for such agents is that they be able to act and system), as a guide to which intentions to adopt and how
proactively to anticipate user needs, opportunities and prob- to pursue them. Finally, we would like to develop a theory
lems, and then act on their own initiative to address them. of proactivity that characterizes both when an agent should
This position paper outlines some initial thoughts on desired take initiative to assist the user, and the nature of the assis-
forms of proactive behavior, and identifies technical chal- tance that should be given. In this position paper, we set out
lenges in developing systems that embody such behaviors. initial thoughts on these topics and identify some important
questions for future research.
Introduction Characterizing Helpful Assistance
We are interested in developing intelligent personal assistive Ethnographic studies of human work habits and task man-
agents that can aid a human in managing and performing agement (e.g., [1, 3]) reveal that people usually achieve all
complex tasks. Our overall goal is to reduce the amount of their important tasks. We become adept at multi-tasking and
effort required by the human to complete the tasks she in- remembering the things that really matter; however, we fail
tends. Effort here encompasses both the activities necessary to achieve perfectly tasks with soft deadlines or forgettable
to perform the tasks, and the cognitive load in managing and details.
monitoring them. Thus, a personalized assistive agent may Let us distinguish tasks performed solely by the user
aid its user directly by performing tasks on her behalf or in (user tasks) from those performed solely by the agent (agent
conjunction with her [7], and indirectly through actions such tasks), and those performed jointly in partnership (shared
as providing context for her work, minimizing interruptions, tasks). We assume that the user enjoys an instrumented
and offering suggestions and reminders [3]. work environment and that she employs electronic artifacts
We are exploring these ideas within a system for intelli- to keep track of her tasks, commitments, and calendar.
gent personalized assistance called CALO [11]. The focus It is not our goal to address the general problem of infer-
for CALO is to support a busy knowledge worker in dealing ring a user's intent [9, 5]. Although research on CALO en-
with the twin problems of information and task overload. compasses recognizing from her actions what task a user is
CALO's current task-related capabilities are grounded in a working on [11], our starting point (for the moment) is that
delegative BDI model [12], in which the system adopts in- the user has entered a description of her user tasks and tasks
tentions only in response to being explicitly assigned them assigned to her CALO into an electronic todo list. In addi-
by the user. CALO can perform a variety of routine of- tion, we assume that CALO is told or can infer a mapping
fice tasks delegated by the user, such as arranging meetings from these entries to formal models within a task ontology
and completing online forms, as well as more open-ended (i.e., the tasks have associated semantic descriptions).
processes such as purchasing equipment or office supplies Within this setting, we can envision a personal assistive
and arranging conference travel. agent aiding its user in many ways:
One limitation within the current CALO framework is the
· achieve a goal/perform a delegated task
lack of a proactive capability that would enable CALO to an-
ticipate needs, opportunities, and problems, and then act on · collect information
its own initiative to address them. We are interested in devel- · share information (with user or team)
oping proactive behaviors along these lines within CALO, to
increase the overall effectiveness of the system as a personal · filter information/reduce interruptions
assistant. · remind and notify
· summarize across projects/time In the above example, CALO's actions are pertinent to the
· provide context for a task or message important upcoming meeting. CALO itself is not capable
of reviewing the paper; identifying a colleague who poten-
· monitor task progress tially is able, CALO does not delegate the task from your
· anticipate potential problems (e.g., lack of resources) todo list automatically, but leaves you in control to take the
suggestion or not. This suggestion and the preparation of
· explain status, reasons, and failures background materials are both safe, in that they result in no
· intervene in a user or shared task to help changes of state, other than a gain in information. Through-
· suggest relevant actions (e.g., link to resources for current out, CALO's actions are unobtrusive: the communication is
task) via a chat message with context, and the completed infor-
mation gathering is again in context, attached to the relevant
· provide team liaison (e.g., manage delegation requests) artifacts in your working environment.
· instruct the user (e.g., how best to use CALO)
· learn (e.g., set an agent learning goal of how to do a task) Demands on a Theory of Proactivity
Example Scenario CALO observes the items in your A theory of proactivity will likely have much in common
electronic todo list, what you are working on currently, what with theories of collaboration, since both are rooted in the
you have delegated to your CALO and to other people, and notion of an agent taking action to assist another. The lead-
your commitments for the week ahead. CALO assesses that ing candidates in collaboration frameworks are Joint Inten-
your workload is likely to be uncomfortably high at the end tion theory [2] and SharedPlans theory [7].
of the week. Via a chat message, CALO offers you a re- Joint Intention theory [2] formalizes the communication
minder of an important meeting early next week, with the acts between agents to establish and maintain joint belief
suggestion that a paper review (on your todo list) could be and intention: the obligations on what "message" to com-
transferred to a colleague (whom CALO identifies as having municate and under what circumstances to do so. Shared-
appropriate expertise and time in his schedule), to leave you Plans theory [7] specifies the collaborative refinement of a
time to focus on the meeting. In addition, CALO begins to partial plan by multiple agents; it handles multiple levels of
prepare background material for the meeting without being action decomposition and partial knowledge of belief and
explicitly asked. It attaches the relevant documents to the intention.
item in your todo list and the event in your calendar. The characterization of how to provide proactive assis-
tance could likely be modeled as a variant on these collabo-
The above scenario illustrates two distinct types of proac- ration theories in which certain of the requirements for mu-
tive behavior for an agent. The first type, which we call tual belief and commitment are relaxed. Characterization of
task-focused proactivity, involves providing assistance for a when to act proactively, however, is not considered within
task that the user either is already performing or is commit- these theories. Here, we consider some of the factors that
ted to performing; assistance takes the form of adopting or bear on this control issue.
enabling some associated subtasks. Task-focused proactiv- A helpful assistive agent weighs the cost and benefits of
ity is exemplified in the above scenario by CALO collecting potential intentions and the plans to achieve them [5]. We
background information in support of a scheduled meeting. first need a theory of action consequences in order to define
The second type, which we call utility-focused proactiv- the concept of a safe action. For example, is a safe action
ity, involves assistance related to helping the user generally, one that maintains world state, other than adding to user or
rather than contributing directly to a specific current task. agent beliefs? Or is an action safe provided the state changes
An example of this type occurs in the scenario when CALO can be reversed (and at what cost)? Does safety also encom-
takes the initiative to recommend transferring a paper review pass not interfering with the user's actions? This theory is
task in response to the detection of high workload levels. required for the agent to assess what actions are safe.
This action is triggered not by a desire to assist with any in- Second, we need a theory of user desires to describe what
dividual task on the user's todo list, but rather in response to are the long- and short-term goals of the user. Such a theory
a higher-level desire (namely, workload balancing). provides a means of assessing the value of each agent action
Principles We identify five principles to guide proac- in terms of the user's objectives. The question for the agent
tive behavior (compare the principles for intelligent mixed- is then: when are (unsafe) actions to be considered? If a task
initiative user interfaces in [8]): has many safe actions and high perceived benefit, should it
be barred because one action is potentially unsafe, such as
· unobtrusive: not interfering with the user's own activities
accepting a meeting request on the user's behalf?
or attention, without warrant
Third, a theoretical basis to support the helpful behavior
· valuable: pertinent to advance the user's interests identified above must account for at least (1) user, shared,
· capable: within the scope of the agent's abilities and agent tasks; (2) acting in support of another agent's
goals; (3) restricting actions to those that are perceived safe.
· safe: without negative consequences Finally, it must admit the timeliness of action and interac-
· user control: exposed to the scrutiny and according to the tion, in order to support the agent's unobtrusive, pertinent,
mandate of the user user-controllable mixed-initiative assistance.
2
Challenges in Ongoing Work helpful and thus trustworthy assistant over time. A part of
personalization that is central to user experience is that of
Acquiring Agent Understandable Tasks The scenario interaction [8]. When and by what modality does the agent
above hinges on the assistive agent's ability to infer associa- communicate its beliefs and actions? When is it better not
tions among and reason over information such as todo items, to act, to interrupt and ask, and to act [10]? What interfaces
calendar entries, projects, resources, plans, current task and provide for efficient communication and collaboration? An
location, and agent capabilities. effective assistive agent will deliberate not only about inten-
How can these various aspects of knowledge be popu- tions to act, but also intentions to communicate.
lated? For example, semantic information about user and
shared goals is critical. Three possible sources of informa- Acknowledgments. The authors thank other members of the
PExA project. This material is based upon work supported by the
tion are (1) inference from user actions (i.e., intent and plan
Defense Advanced Research Projects Agency (DARPA), through
recognition) [9]; (2) explicit user declaration in a semanti-
the Department of the Interior, NBC, Acquisition Services Divi-
cally grounded manner (as we assumed above); (3) inference
sion, under Contract No. NBCHD030010.
from user non-semantic statement, possibly confirmed with
an explicit disambiguation request. In the third case, for ex-
ample, the agent might link the informally specified entry References
"book conference" on the user's todo list to the known task [1] V. Bellotti, B. Dalal, N. Good, P. Flynn, D. G. Bobrow, and
conference-travel by drawing on techniques such as N. Ducheneaut. What a To-Do: studies of task management
those described in [6]. towards the design of a personal task list manager. In Proc.
of CHI'04, pages 735742, Vienna, Austria, 2004.
Since studies show that people decompose their work into
projects and todos at differing abstract levels [3], a related [2] P. R. Cohen and H. J. Levesque. Intention is choice with
commitment. Artificial Intelligence, 42:213261, 1990.
challenge is the levels of abstraction at which to define a
task ontology, and how to relate user-specified tasks into it. [3] M. Czerwinski, E. Horvitz, and S. Wilhite. A diary study of
task switching and interruptions. In Proc. of CHI'04, pages
While task-focused proactivity seeks to provide assistance 175182, Vienna, Austria, 2004.
to the user with immediate, tangible goals, in contrast utility-
focused proactivity addresses more general objectives of the [4] J. L. Edwards and G. Scott. LOCATE intelligent systems
user. Utility-focused proactivity requires a representation of demonstration: Adapting help to the cognitive styles of users.
In Proc. of AAAI-06, pages 19371938, Boston, MA, 2006.
user desires that captures the user's unstated interest goals
(in the terminology of the OCC cognitive model [13]) as [5] A. Fern, S. Natarajan, K. Judah, and P. Tadepalli. A decision-
well as explicitly stated achieve and replenishment goals. theoretic model of assistance. In Proc. of AAAI'06 Workshop
Since interest goals differ between people, a helpful assis- on Modeling Others from Observations, Boston, MA, 2006.
tive agent requires such a model of its user, in order to assess [6] Y. Gil and T. Chklovski. Enhancing interaction with to-
the value of agent actions. do lists: Using artificial assistants. In Proc. of AAAI 2007
Spring Symposium on Interaction Challenges for Intelligent
Acquiring Task Parameters To act on an intention to per- Assistants, Stanford, CA, 2007.
form a task, the agent must have instantiations for the task's
[7] B. Grosz and S. Kraus. Collaborative plans for complex
input parameters. For example, it is not enough to iden- group action. Artificial Intelligence, 86(2):269357, 1996.
tify a conference-travel task: the agent must know
to which conference the user is referring. One approach is to [8] E. Horvitz. Principles of mixed-initiative user interfaces. In
acquire the parameters to instantiate the task fully, by asking Proc. of CHI-99, pages 159166, Pittsburgh, PA, 1999.
the user to specify them. Since this risks disturbing the user [9] E. Horvitz, J. Breese, D. Heckerman, D. Hovel, and K. Rom-
with a perceived irrelevant request (unless the user has just melse. The Lumiere project: Bayesian user modeling for in-
asked her agent to perform the task, the request comes out of ferring the goals and needs of software users. In Proc. of
the blue), another approach is to guess the parameter values UAI'98, pages 256266, Madison, WI, 1998.
from learned history, or to perform information-gathering [10] E. Horvitz, C. M. Kadie, T. Paek, and D. Hovel. Models of
actions (e.g., look at the user's calendar). attention in computing and communications: From principles
to applications. Communications of the ACM, 46(3):5259,
A second approach is to act on a set of possibly matching, 2003.
partially-instantiated tasks by performing safe, conditional
actions, one for each possible task instantiation. For exam- [11] K. Myers, P. Berry, J. Blythe, K. Conley, M. Gervasio,
ple, gather flight and hotel quotes for each conference. A D. McGuinness, D. Morley, A. Pfeffer, M. Pollack, and
M. Tambe. An intelligent personal assistant for task and time
strong notion of safety is needed here, since the agent must management. AI Magazine, 28, 2007. To appear.
not reserve a flight to each venue! In conjunction or instead,
the agent can perform conformant actions that support any [12] K. L. Myers and N. Yorke-Smith. A cognitive framework
of the possible tasks. For example, whatever the destination, for delegation to an assistive user agent. In Proc. of AAAI
2005 Fall Symposium on Mixed-Initiative Problem-Solving
the user will need to submit a travel authorization form, and Assistants, pages 9499, Arlington, VA, 2005.
many of the fields can be prepopulated.
[13] A. Ortony, G. L. Clore, and A. Collins. The Cognitive Struc-
Personalization and User Interaction By adapting to its ture of Emotions. Cambridge University Press, 1988.
user's preferred working and communication styles, and her
capabilities and experience [4], an agent becomes a more
3