William A. Adams1
1Adjunct Associate Professor of Psychology, Chapman University College, Bangor Learning Site, PO Box 2120, Silverdale, WA 98383, USA.adamswa@earthlink.net
Abstract. This paper suggests a psychological approach to the problem of context management in intelligent software. The user’s motivational context is taken by the agent as its own motivational context. While traditional task-specific context problems still must be dealt with, the agent can use its semi-autonomous motivation to navigate these problems on the user’s behalf. An implementation strategy is suggested.
1 Introduction
AI applications are notoriously brittle, unable to adapt when context changes and often unable to discriminate appropriate from inappropriate context for a task. But people are remarkably adaptable, possibly because humans maintain two different kinds of context, while AI applications have only one. Humans make contextual assumptions about their physical environment, but also have the ongoing context of their own motivation. The motivational context subsidizes the other one. It is this linked, dual-context system that makes humans resilient under conditions of change. The paper defines the two kinds of context then suggest how an AI application might incorporate them both.
1.1 Two Kinds of Context
When context changes, the validity of descriptions and the truth of assertions about a situation may be lost. Who has not been startled when doing a word-processing task to see out of context results. If you want to change instances of the word "as" to "like," you do not expect the software to change "Christmas" to "Christmlike." We try to control such errors by explicitly giving contextual detail in our instructions to the word processor. We tell it, "Find whole words only." That is an example of articulating the context for an environmental object. We can call that kind of context "environmental context."
Another kind of context is "motivational context." Consider a help system. When the user clicks the help button, she gets reference material concerning the operation she was attempting to perform when she interrupted herself to click the help button. When such systems work well, they deduce the user’s motivational context -- what the user was trying to do. If she is in the middle of resetting page margins when she clicks the help icon, the help system takes her query to be about margins. But if the user actually wanted help concerning printing, even though she is currently resetting page margins, the help system fails to be sensitive to her motivation. The help system operates on an inference about the user’s motivation, not on knowledge of her actual motivation.
1.2 Relationship Between Motivational and Environmental Context
For a human, the relationship between an object and its purpose is not neat. Because human purposes are indefinitely variable, a cup can be a paperweight if that is one’s wish, and a coin can be a screwdriver. There is no way to adequately represent a cup and a coin as elements of an environment without knowing something about the motivational context of the person who might use them. Searle[1] handles this relationship between objects and motivation by identifying a class of "socially constructed objects." A screwdriver used as a doorstop is still really a screwdriver, he says, while as-a-doorstop it is merely a "socially constructed" object. That formula isn’t much help for designing AI systems since we cannot tag every environmental object for an unbounded set of potential uses. Instead we need to be more clear about the motivational context which determines the environmental context of objects, for each person, in each situation.
The relationship between motivational and environmental context constituents is paradoxically one-to-many. A person has many motives, and the environment has many features. But with respect to any one dominant motive, its change can undermine the entire context of all the environmental features. If I move from my computer keyboard to my piano keyboard, the entire context of the internet becomes instantly irrelevant to me, and even the term "keyboard" takes on a new meaning. Any set of common-sense facts, no matter how extensive, can be rendered irrelevant by a single change in motivational context. On the other hand, a single change in an important environmental object can stymie a large swathe of one’s motivation and paralyze all action. This happens, for example, when a person suffers a sudden unexpected loss. The context problem is actually a problem of alignment between motivation and the objects that satisfy it. That is why no amount of elaboration of environmental context alone can solve "the context problem" of AI applications.
2 Motivation and the Frame Problem
Minksy[2] gives this exemplar AI context problem: "Fred told the waiter he wanted some chips" (p. 68). Minsky lists a dozen or so inferences, adapted from Lenat [3], that can be made from that sentence, such as "Fred wants potato chips, not wood chips, cow chips, or bone chips." This is followed by a generalized hierarchical "Architecture of Representations" which describes "levels" of detail from which these inferences might be drawn. Minsky proposes that the problem of changed environmental context can be managed by representing the environment in ever larger, more inclusive, more abstract propositions.
There are two problems with this approach. First, there is no limit to the detail that might be needed, and no sides to the breadth of interconnections among details. Minsky, like everybody else, must draw some line that separates germane environmental detail from the irrelevant or unimportant. After that, any environmental object "below the line" breaks the context. But unimportant to whom? For what purpose? That leads to the second problem.
It is not hard to imagine a fellow named Fred pursuing industrial espionage who asks his co-conspirator restaurant employee for stolen microprocessor chips. Which is the more likely inference from Minsky’s sentence: that Fred is hungry or that Fred is a thief? It depends on who Fred is and what he’s up to. An elaborate hierarchy of frame-based environmental context is no help. Understanding of the environmental context depends entirely on Fred’s motivation.
We need to know, who is Fred, and what is his motivation? Does Minsky know? Probably not. "Fred" is introduced as a hypothetical case, not a person, and Minsky’s test sentence is not a real assertion of fact, only an example of an assertion. Minsky pulled Fred out of his hat. If Minsky had given a real anecdote about a real friend, Fred, who was hungry and ordered some potato chips, then we might understand Fred’s motivation better and establish some basis for understanding his environmental context.
If Minsky’s test sentence were seen as part of a formal system rather than as an example of natural language, we would have no argument. In some formal system of language, perhaps the word, "waiter" constrains "chips" to the category "food products," so we could infer that "chips" refers to potato (or some food) chips. But natural language understanding is not based on a closed system like the rules of chess and the polysemy of "chips," can only be disambiguated by reference to situational context. Minsky assumes, or pretends, that the sentence is a real natural language sentence, an assertion about the world which could be interpreted by an intelligent agent. He suggests that an agent could infer from the sentence such things as, "Fred and the waiter speak the same language. Fred and the waiter are both human beings. Fred is old enough to talk (2+ years of age). The waiter is old enough to work (4+ years, probably 15+)." (pp. 68-69).
But Minsky cannot legitimately infer that Fred is old enough to talk unless he already assumes the validity of the entire banal restaurant scenario he is attempting to "infer." Fred could be speech-impaired, communicating only by sign language that a waiter could not understand. Fred could be delusional, unable to make his real intentions clear, speaking of "chips" when he really wanted fish. Fred could be alone, hallucinating in his bathtub. Fred could be a dog wagging his tail, his "request" apparent only to his owner.
Minsky uses a common name, "Fred," which readily invokes the image of an ordinary guy that we can imagine to be much like ourself. Minsky refers to "the waiter," using the definite article to suggest that we already know what is being referred to: not a person waiting at a bus stop, but a person fulfilling the socially designated role of restaurant waiter. Thus the example invokes an imaginary frame much larger than the explicit representation, in which we assume that Fred is a person just like us. If we assume a typical environmental frame as context, we also assume an implicit parallel frame of motivation for the actors. We then feel we can say, "I know what I would intend in that situation, therefore I know what Fred intends."
But isn’t the whole point of the context problem the fact that this approach produces errors? In fact we don’t know much about other people’s motivation nor really, about our own. Minsky has seduced us into believing we understand far more about the situation than we actually do and therefore that we could program an AI agent to understand as much. But we cannot effectively represent an environment to our agent without an analysis of the agent’s motivation.
3 An Intelligent Agent is Motivated
If an AI program is to behave intelligently in the face of changed or ambiguous context, perhaps it should draw upon similar resources to those that a human uses to deal with context. An intelligent agent would have an explicit motivational context as a person does. The obvious model for and source of motivation for the agent is the user. The AI agent must establish and maintain a social relationship with the user, not for the sake of "user-friendliness" but so that the agent can understand and absorb the user’s motivation as its own.
This reasoning implies that as designers of intelligent agents, we need to revise our conceptualization of what the agent is. We tend to think of the agent only as a machine, an algorithm. We might add user-friendly features to the machine to facilitate the user’s comfort. But instead of adding a superficial anthropomorphic interface to the machine, we need to conceptualize the agent as a motivated entity from the start. As the user’s literal agent, the software agent should be motivated as the user is. The agent would make decisions based on both its own mental state and its understanding of the mental state of the user [4].
If we were to write intelligent agents as motivated, social beings, the user’s relationship to the agent would not be one of machine operator to device, as it is today. Rather, it would be a social relationship, like any based on mutual consideration. Instead of operating a "search engine," for example, the user would turn to an agent for help in finding something. The user would ask the agent to adopt her goal as its own [5]. It would be a different mental attitude towards agents than what we are accustomed to now.
3.1 Example of a Motivated Agent
If the user is trying to find and buy a consumer product, she might ask an agent to search for her. In a typical search scenario, the user might fill out a form describing the category of item ("electronics"), the type ("CD Writer") and the price ("under $300"). The list of search results could be quite large. Using a socially engaged and motivated agent, the scenario would look different.
|
AGENT: |
Why do you want it? |
|
USER: |
I’d like to be able to play music as I work on my computer. |
|
AGENT: |
Why not turn on the radio? |
|
USER: |
I want to hear my own music, and with a writer, I could save downloaded music files and play them later. |
|
AGENT: |
Have you downloaded and played digital music before? |
|
USER: |
No, but I understand it is quite easy to do. |
|
AGENT: |
Would you use the CD writer at home or at work? |
|
USER: |
Probably both, if it was easy to carry and connect. |
|
AGENT: |
Would the music disturb people around you at work? |
|
USER: |
We take turns putting our choice of music on the player in the workgroup, so it’s always loud. |
The interview could continue. Based on it, the agent might retrieve a basket of goods: a CD writer with an appropriate interface for the user’s home equipment, software that reads and writes music files for random access, and which can write standard format CD-R for playing as group music at work; several blank CD-RW and CD-R discs and a labeling system. It seems likely that only playing, not writing CDs would take place at work (the user could be queried on this). In that case, portability of the device would not be a high priority. Instructions, either a book or an URL, would be suggested for how to download and play music files from the net. Auxiliary speakers for the home computer might be recommended, and a CD carrying case.
In this example, the agent attempts to understand what the user is trying to accomplish, not just what the user is trying to buy. Therefore the agent is less likely to return choices that are irrelevant to the user, such as automobile-mounted CD devices. The agent is more likely to return related items, such as a CD carrying case for taking CDs from home to work, or an MP3 player with earphones for work. On the other hand, if the user were interested in converting a vinyl music collection to CD, the agent would look at digital sampling equipment to match the CD-writer. This agent has three categories of knowledge to work with: (a) plenty of task-specific knowledge about CD equipment and related goods, (b) knowledge of the user’s situation, including home and work computer configurations, and (c) understanding of the user’s motivation. It is the last category, knowledge of the user’s motivation, that makes the agent able to deal with context issues that might arise.
The intelligent agent tries to discern what the user wants, rather than simply fetching items that match a literal request. For example, the agent is not confounded by finding a software package called a CD-writer that generates marketing copy for a bank’s certificates of deposit. If there was any doubt, the agent would simply ask the user for clarification. By taking on the user’s motivation as its own, the agent’s behavior is grounded across changing environmental contexts.
3.2 Motivation Vs User-Interface Design
Many software programs query the user’s needs prior to acting. Tax preparation programs interview the user extensively prior to writing a return for example. Such programs are successful within their extremely limited domain by mapping constrained user choices to defined parameter selections. But this is more a user interface issue rather than an example of motivated software because the interaction is narrowly specific to the software’s immediate task.
A broader scope is seen in the anthropomorphic cartoon interface for Microsoft’sTM recent PC operating system. "Victor," the help agent, guides the user through any number of software tasks. This approach is effective because the user’s motivation is assumed to be precisely the correct operation of that particular software. The user’s motivational context is inferred from the encapsulated environmental context. Victor never asks questions about the user’s intent. The user’s actual goal, perhaps to persuade a committee, extends far beyond Victor’s capacity to address. Users understand that Victor is an interface and that they must adapt to it, not vice-versa.
Another approach to motivational context tries to infer what the user would like, based on recent behavior. An online retailer infers from yesterday’s purchase of a book on the history of jazz that I might want to buy a collection of recorded jazz. But that inference may be irrelevant today. A motivated agent would work for me, not for the retailer, and would know that today I am researching Alan Turing, not jazz. And I would not expect to be advised, "Click here for great prices on Alan Turing." Current attempts to provide motivated agents suffer from narrow scope, both in subject-matter domain and in time. They are still only kinds of user interface. A properly motivated agent has an ongoing relationship with the user which grounds environmental context decisions over time and over topics.
4 Implementation Ideas
In a simple strategy for a well-motivated agent, a program would ascertain and adopt the user’s motivation as its own, something a person cannot do. A person’s fundamental motivation is intrinsic, although goal-oriented expression of motivation is acquired. A person or an agent can adopt another’s goals as his own, but goals are the objects of motivation, not the sources of them. The motivated agent adopts the sources of the user’s motivation, becoming a semi-autonomous motivational clone of the user with respect to the objects in the environment.
A series of dialog boxes could determine the user’s motivation and establish for the user the agent’s capacities. Such a transaction could establish a relationship of mutual expectation and trust between the two: for the user, literally, and for the agent, a representation of the relationship based on a motivational profile of the user. The dialog must be general and recurrent rather than singular and task-specific. The dialog would aim to make explicit needs and expectations. Appropriate questions of the user might include, "What things interest you? What processes suit you? What are your long and short term goals? What will success on this task look like for you? What would make you mad? The agent essentially interviews the user as if it were considering the user for employment.
It is important to note that the interface need not be natural language or even very sophisticated. The user might select germane questions from lists and type in answers as words and phrases that the agent can maintain in lists. The point of the transaction is not to create conversation for the sake of conversation, but for the agent to establish a model of the user’s motivation.
Task-specific parameters still must be set for the environmental context, but the agent can embed these in the context of the user’s motivation. The user has short-term, task-specific intentions, a medium-term ("project-sized") motivational context and a long-term (personal development) motivational context. These are the sources of the user’s motivation. The agent also garners motivational data from various user requests over time. The model of the user’s motivation becomes the agent’s own motivational context, a relationship that defines a perfect agent.
4.1 Representation of Motivational Context
There are many ways the agent might represent the user’s motivation. For example, within a temporal framework (short, medium, and long-term), the agent could classify the user’s bodily interests (e.g., in comfort, health, beauty, skills, tools, strength), social interests (e.g., in love, power, status, family, things and information, community, affiliation, achievement), or philosophical/intellectual interests (e.g., spiritual, scientific, literary, self-help). That particular tripartite division of motivation is adapted from the ancient Greek theory of education, but other schemas are possible. The sophistication of the motivational schema determines the ultimate effectiveness of the agent. It is an abstract taxonomy of the user’s motivation, not a classification of goods and services that the user might want. The agent adopts that representation as its own motivation.
4.2 Object Tags
How does the agent map the motivational context to the environmental context of a specific task? Each environmental object that might be encountered or retrieved by the agent must be tagged with a set of motivational identifiers that can connect to the user’s and agent’s motive taxonomy. Since the taxonomy is much broader than any specific task, the object tags are also at a high level of generality. The motive tags are in addition to the ordinary feature list that defines a retrievable item today. The motive tags allow the agent’s search of the environmental context to be guided by the fit between the motivational profile and objects’ motive tags (as well as the traditional fit between object features and the agent’s domain knowledge).
Object motive tags can come from two sources. The most direct source is the user herself. For example, the user could execute a rating scale on an object or class of objects. It would ask for reactions to or impressions of the object, using descriptors that can map to any motivational profile. For example the user might indicate if an item is associated more with strength or with weakness, value or triviality, beauty or plainness, etc. The resulting user reaction profile for the object, similar to a semantic differential [6], constitutes the set of motivational tags for an object. This "reaction" method addresses the fact that most people do not articulate their own motivation very well. A well designed reaction profile will be general; suitable for a large range of objects and users, yet will map determinately to the agent’s representation of its particular user’s motivation. The user could perform such reaction profile ratings for the key terms of a particular search request, or for entire categories of items, such as all the categories presented at a particular level of a web directory.
Ideally, the motivational tags on each object would have been defined by whoever offers the object for retrieval, according to some standard set of categories. While less specific to a user’s transient reactions, this would characterize the specific kinds of satisfaction offered by the object for specific kinds of consumers. The consumer would have considerable power of active selection based on their own motivation, expressed through their agent.
Conclusion
An agent is someone who acts in my behalf. If all I need is to pick up a pencil I dropped on the floor, I do not need a very intelligent agent to fetch it for me. But if my needs are more nebulous and the objects which would satisfy them are unknown to me, an intelligent agent is just what I want. Such an agent must understand my motivation, for even I do not know exactly what I want. "Understanding" then means ‘able to connect my motivation to a set of objects which could satisfy it.’ This is the rationale from the user perspective for motivated software agents.
From the engineering perspective, "the context problem" arises when one or more objects has been represented in terms of a particular environmental context when a different context actually prevails. Rather than try to map all possible environmental context frames to each other, we might take a cue from human psychology, where the context of a person’s motivation is her portable disambiguator of environmentally bound object definitions. The idea of a motivated software agent suggests a different way of dealing with the context problem.
References
1. Searle, J.R. (1995). The Construction of Social Reality. NY: The Free Press.
2. Minsky, M. (2000). Commonsense-based interfaces. Communications of the ACM, Vol. 43, No. 8, pp. 67-73.
3. Lenat, D.B. (1990). The Dimensions of Context-space. www.cyc.com/publications.htm.
4. Rosis, F., Covino, E., Falcone, R., & Castelfranchi, C. (1999). Bayesian cognitive diagnosis in believable multiagent systems. International Workshop on Belief Revision, Trento. http://aos2.uniba.it.8080:/papers/cogdia.ps.
5. Conte, R., & Castelfranchi, C (1995). Cognitive and Social Action. London: UCL Press.
6. Osgood, C.E., Suci, G.J., & Tannenbaum, P.H. (1957). The Measurement of Meaning. Urbana, Ill.: University of Illinois Press.