By: David Dattner (934265)

Cpsc 547 - Intelligent Agents Class Presentation

Partners, Graham Neumann & Chris Marta.


More general information about Personal Software Assistants

Personal software assistants help users find information in databases, schedule calendars, or manage work flow will need significant customization for each individual user. An example, from the readings, is an assistant that helps schedule a user's calendar will have to know that user's scheduling preferences.

- CAP (Calendar Apprentice)

An example of this is CAP(Calendar Apprentice), that learns it's users scheduling preferences from experience. With CAP results are summarized from about 5 years of experience, during which it has learned several thousand rules that characterize scheduling preferences for it's different users. It is not difficult to imagine a future where assistants will operate as software secretary, providing service at work and at home, such as paying bills, making travel arrangements, and locating information from electronic libraries. Just as the success of a real secretary depends on the his or her knowledge of the employer's particular habits and goals so does the agent's success dependent on these abilities to learn and adapt to the user. A perfectly customizable software assistant which will work for any individual is not a simple task to create.

- explicit customizability vs. auto-learning

Many programs provide simple parameters that allow users to customize behavior explicitly. We see this quite often with text editors which allow users to set default type fonts and directories etc... Yet this type of approach is limited, in that continual updating is tiresome and detailed.. Usually, the more sophisticated an assistant the less the users needs to program their preferences explicitly. Instead the software assistant should learn through experience, just as humans do.

COACH (A Teaching Agent that Learns)

Cognitive Adaptive Computer Help, is a system that records user experience to create personalized user help. For example if a user is learning a new operating system or application COACH watches the user's action to build an adaptive user model (AUM) which chooses the appropriate advice. It is important to note that just like a coach on a football team, the COACH does not interfere with the user's action but instead only supplies advice along the way. Moreover, it would also decide on which particular method to convey it's information, hence using description, example, syntax, timing, topic, style and level of help according to the user's experience and proficiency.

An example: LISP Implementation COACH

Developed by Dr. Ted Selker at IBM's Almaden Research Center in California, this teaching assistant holds programming students by the hand as they learn the intricacies of Lisp. In a test using IBM programmers with no Lisp experience, student/Coach teams completed five times more training exercises involving database function calls than did students who didn't have the aid of an electronic tutor. At the end of the training sessions, most Coach pupils said they liked Lisp; their untutored counterparts weren't so enthusiastic. Skeptics may scoff at agent hype--and the dearth of commercial programs--but if an agent can actually make using Lisp an enjoyable experience, maybe anything is possible. Selker, who is manager of USER (User Systems Ergonomics Research) at the IBM center, has been evolving Coach along these lines over the past decade. Coach, which is being converted from Lisp to C++, contains three independent but related sets of knowledge to do its work. The first is its user model, where the program builds and maintains information about a user's Lisp-programming proficiency, his or her mistakes (and the chosen fixes), and what worked and what didn't work in terms of coaching. Coach also tracks the rate of change in each user's proficiency (and whether it's getting better or worse), and for each learnable task, how long it's been since it was last used. If a user has done something correctly in the past and has trouble with it now, Coach can present an example from the user's own code.

- User Model

The user model is persistent--it remains associated with a user from session to session. In a teaching environment, the user model could be made available to a teacher to help him or her understand just where a student might be having difficulties. In a commercial implementation, the coach agent might pass the user model onto a human customer-service representative if the agent determines a customer's problems are beyond the agent's capabilities.

- Domain Knowledge

Second, Coach has knowledge about its subject matter. In the case of Lisp programming, Coach knows Lisp syntax, library functions, and concepts (e.g., evaluation, iteration, and recursion). This knowledge base grows over time, automatically incorporating user-defined functions into its repertoire. Such a facility would be a welcome addition to any programming environment, particularly in a team setting in which you're writing code for and using code from other programmers. Your agent might suggest using an existing function or object before you reinvent one yourself. Keeping the domain knowledge separate from knowledge about coaching has made it easy to apply Coach's framework to new domains. For example, Selker says Coach helped a summer intern inexperienced in programming create a Unix help system in just 10 weeks.

- Coaching Rules

The third knowledge set consists of coaching rules that tie user knowledge and the domain together. These rules help Coach gauge a user's level of experience. Update rules determine when the program should refresh the user model to indicate that he or she has mastered a problem or when it's appropriate to present more advanced usage options for a particular feature. Consistency rules make sure that the user model doesn't contradict itself when the model is applied to related subjects. Finally, presentation rules determine how help will be presented to the user. For example, a user who's just starting out would want basic information, while an out-of-practice user may only need a reminder.

Agents (Programs that Simulate a Human Relationship)

The job of the agent is to take the place of another experienced human user and teach. This means that the agent should perform tasks for the user that the user does not know how to do yet, with emphasis on the fact the user will then learn how to perform this task. One of the main concerns with agents is the creation of a user dependency on the agent. In other words the agent should be used much like a coach in that it should offer advice but should not be the sole voice and dependency of the user. How Might People Interact With Agents Many people fear the new technology of agents. Many fictitious and extravagant claimed have been as seen in the example of 2001 a Space Odyssey with Hal. This view of the agent especially when modified by the term intelligent brings forth images of human like computers working without supervision on tasks thought to be for our benefit but not necessarily to our liking, Yet some actually believe that the main difficulties with the development of intelligent agents will be social and not technical. The question then becomes how will intelligent agents interact with people and how might people think about agents? For people to feel comfortable with automatic, autonomous actions some factors that must be considered are:

Factors that must be met to ensure users feel comfortable with agents:

-ensuring that people feel in control of their computational systems

- the nature of the human-agent interaction

- built in safe guards to prevent run away computation

- providing accurate expectations and minimizing false hopes

- privacy concerns (a subset of the feeling of control problem)

- hiding complexity while simultaneously revealing the underlying operation

The Role of Emotion in Believable Agents

Many will argue that for a person to put faith in the ability of an electronic agent, the agent must exhibit human characteristics and qualities. Providing the illusion of life. AI researchers are attempting to find these essential qualities of humanity. In the article, artists (animators) and AI scientists were both questioned as to what traits they thought believable characters should have. The qualities the animators judged important are not identical to those that AI researchers emphasized, though there was some overlap. Emotion was deemed very important by the animators. They argued that an agent which is not emotional gives the feeling of indifference and lack of caring and motivation.

- Case study with AI scientists and Disney animators

According to Thomas and Johnston, of Disney, to portray emotion the animator must remember several key points:

-The emotional state of the character must be clearly defined.

-The thought process reveals the feeling.

-Accentuate the emotion.

-Use time wisely to establish the emotion, to convey it to viewers, and let them savor the situation.

Trust

Building and maintianing trust between user and agent requires several issues to be addressed:

- User knowledge and Domain knowledge

Two components of user knowledge are included in our model:
  • domain knowledge
  • technical knowledge. Both domain and technical knowledge of intelligent agents impact the user's

    ability to predict the agent's actions and assess the agent's technical compentance.

    - Predictability

    Predictability is based on consistency (Rempel, et.al., 1985). It is reasonable for the user to expect the agent to perform consistently on frequent tasks. When so, the agent is more predictable, which has a positive impact on trust. Ensuring that users are comfortable with agents' changing behavior is important. Designers of intelligent agents need to keep this in mind when deciding upon how much of the agent's actions should be visible to the user.

    - Dependability

    As a relationship grows, people judge the characteristics of another, rather than specific actions. One of these characteristics is dependability. Predictability provides evidence of dependability, but goes deeper. Dependability deals with a specific In the context of intelligent agents, ideally users are confident that agents will step in when their help is needed. If they are not, they will spend more time monitoring, or performing tasks themselves, making the user - agent relationship less effective.

    - Technical Competence

    For a successful user - agent relationship, users must feel certain that their agents are technically competent of performing assigned tasks. Otherwise, trust will suffer and monitoring will increase, or use will decrease. There are several ways users can judge technical competence. Frequent interaction with the agent can demonstrate the agent's competence, or lack thereof. Early in the relationship extra monitoring may be necessary, until the agent "proves" itself.

    - Fiduciary Responsibility

    Fiduciary responsibility comes into play when an individual lacks sufficient knowledge to judge the actions taken on their behalf by others (Barber, 1980). The usual relationship between an attorney and a client is an example where fiduciary response in a user - agent relationship, the user (e.g. a novice) may not always be in a position to accurately assess the appropriateness of the agent's actions. Assessments of fiduciary responsibility will often be made based on the user's feelings towards Monitoring Behavior One of the judgments individuals must make in client - agent relationships is the proper level of monitoring. In general, proper monitoring helps ensure the efficiency of the relationship. Trust affects this monitoring behavior. Frequency of Use It seems reasonable that a user's assessment of an intelligent agent's trustworthiness impacts how frequently a user will employ the agent. This assumes that the user can choose to use the agent on some circumstances, and not in others.


    References

    Articles

  • Bates, Joseph. "The Role of Emotion in Believable Agents" Communications of the ACM, 122, 7:122-125
  • Norman, Donald A. "How Might People Interact With Agents" Communications of the ACM, 68, 7:68-71
  • Selker, Ted. "COACH: A Teaching Agent that Learns" Communications of the ACM, 92, 7:92-99

    Email Address

    You can mail any comments or suggestions to
    dattner@cpsc.ucalgary.ca
    neumanng@cpsc.ucalgary.ca