Plans and Situated Actions
by Suchman, Lucy A.
Which came first, the action or the plan? The plan, you probably say even without a moments hesitation. This is, however, according to Suchman (1987) a poor way of understanding what really happens when a person sets out to do something. She says that it is only when we have to account for our actions that we fit them into the framework of a plan. Actions are to a great extent linked to the specific situation at hand and are therefore hard to predict by using generic rules. Action, as well as learning, understanding and remembering, is situated.
Suchman criticises the way we sometimes speak, and think, of computers as participants in interactions on equal terms. This is misleading since computers, although they one day might not be, are well behind us in the reasoning department and because they have very limited perceptive abilities.
Human activity can not be described sufficiently beforehand and computers need these plans since they can not properly interact. This is the dilemma investigated in this book. Compared to other forms of skill acquisition, computer based help systems resemble written instructions, which are generic and disassociated from the situation, much more than face to face instructions which are context sensitive and generally more powerful but where the effort has to be repeated.
Attempts have been made to conquer these problems by letting the computerised coaches use tutoring techniques similar to ones used by human coaches. Suchman mentions two systems, WEST (Burton and Brown, 1982) and ELIZA (Weizenbaum, 1983, p. 23). WEST is an artificial coach to an arithmetic game called "How the West was Won". It operates by using a ruled-based system to determine how to coach the person playing the game. A rule can for example be to only give an alternative if it is considerably better, or never to coach on consecutive moves and so on. ELIZA is the collective name of a group of programs made to study natural language conversation between man and machine. The most famous of these programs is DOCTOR which is a program that tries to simulate a psychotherapist. Here the method used was to say as little as possible and thus let the patient interpret the output to mean something that makes sense in view of the patients situation.
Even though these systems show progress in the field of man-machine communication they lack certain abilities that are essential to communication. Because of the situated nature of action, communication must include both an awareness of the local context and a mechanism to solve problems in understanding.
All AI related action research has assumed that the plan has an existence prior to and independent of the action, that it actually determines the action. Intentions are viewed as the uncompleted parts of a plan that is already being executed. This assumption fails to account for intentions to act that are never realised and intentional action for which no plan was formed in advance. In fact, communication primarily effects the models that speakers and hearers maintain of each other according to Cohen and Perrault (1979, p. 179).
The reason for the limitations of these systems is to be found in the theoretical assumptions that of the designs. The planning model states that the significance of action is derived from plans. Therefore the problem of interactions is to recognise and co-ordinate plans. Plans or intentions are understood by the usage of conventions for their usage. This introduces the problem of shared background knowledge. It is not enough to be aware of the local context. There has to exist a wider platform of common knowledge that explains individual actions social meaning.
The solution of the context problem has for the cognitive scientists been to build models of the world. These models have proven reasonably adequate within limited domains such as e.g. medicine but all models taken together still does not at all cover a normal persons knowledge of the world. There seems to be a lot of knowledge, often referred to as common knowledge that does not fit into any model. This problem has so far not been solved by cognitive science and poses great restrictions to the usability of the other models.
Another argument against the plan notion is that the view that background assumptions are part of the actor's mental state prior to the action seems unrealistic. In a conversation for example there would be almost impossible to describe what two persons were talking about without making real-time interpretations. The background assumptions are generated during the course of the interaction.
Suchman calls her remedy to the above described problems situated action. It should be seen as a research programme rather than an accomplished theory. By using a name similar to purposeful action she indicates that it is a reformulation of this theory. Plans are still viewed as an important resource but the emphasis on their importance is considerably weaker than in the original theory. The theoretical base for this reformulation is to be found in a branch of sociology called ethnomethodology.
According to Suchman, plans are representations of situated actions that only occurs when otherwise transparent activity becomes in some way problematic. The objectivity of the situations of our action is achieved through the indexicality of language. By saying that language is indexical Suchman means that the meaning of its expressions is conditional on the situation of their use. At least the communicative significance is always dependent on the situation. Language is a form of situated action. The relation of language to particular situations parallels the relation of instructions to situated action. As a consequence of the indexicality of language, mutual intelligibility is achieved on each occasion of interaction with reference to situation particulars, rather than being established once and for all by a stable body of shared meanings.
Instead of looking for a structure that is invariant across situations we should try to understand how we manage to interact in ever changing contextual settings and still interpret and understand meaning and rationality. The communicative resources used for this include turn taking, adjacent pairs and agendas. Turn taking means that we understand conversations not just by what is said but in what order it is said a question is followed by an answer and so on. Adjacent pairs is an extension to turn taking that denotes e.g. recursively embedded follow-up questions. The turntypes can be pre-allocated as for instance in courtrooms. Agendas is the term for various pre-conceptions of the form and purpose of conversation brought on by its setting.
Suchman has studied an expert help system that regulates the user interface of a copying machine to investigate the problem of the machine's recognition of the user's problems. Data used in the study consisted of videotapes of first-time users of the system. The copier was designed on the assumption that the user's purpose serve as a sufficient context for the interpretation of her actions. The machine tries to use any action from the user detectable to the machine to guess the user's plan and then use that plan as the context when interpreting the user's further actions. The aim of this design was to combine the portability of non-interactive instructions with interaction. The problem is that the relation between intention and action is weak due to the diffuse and tacit nature of intentions.
The study disclosed a serious inability of the machine in reacting properly to input. Human action repeatedly strayed form the anticipated plan. When "Help" meant "What happened?" or "How do I do this?" it was interpreted as "Describe the options of this display." or "Why should I do this?", and so on. The users also frequently misinterpreted the behaviour of the machine since they tried to impose conventions of human interaction in understanding the machine. Suchman divides the interaction problems into two groups conditional relevance of response, e.g. the ambiguity between iteration and repair, and communicative breakdowns. These breakdowns are divided into the false alarm and the garden path. The first term designates the situations where the user is lead to believe that an error has been made when it actually has not and the other means that the user has made an error without noticing it. The system has no ability to discover any of these situations.
This analysis ties the particular problem of designing a machine that responds appropriately to the actions of a user to the general problem of understanding the intentions of purposeful action. From this Suchman extracts three problems for the design of interactive machines. The problem of how to extend the access of the machine to the actions and circumstances of the user, how to make clear to the user the limits on the machine's access to basic interactional resources and how to find ways of compensating for the machine's lack of access to the user's situation.
Instead of using a static model of the user when the system is designed the system needs a mechanism for real-time user modelling that knows when to assist and what to say. This mechanism should be designed based on the following strategies. Diagnosis based on differential modelling, meaning that you use the difference between an ideal (expert) usage of the system and the actual usage to estimate the skill level of the user. When the difference between the developing model of the user and the user's actions gets to big some method for finding the reason should be employed. There should be a division of local and global interpretation of the user where the global accumulation of actions is used to identify weaknesses and misunderstandings. If the user has enough information to identify and repair errors it is considered to be constructive problems. The system should transform non-constructive trouble into constructive.
Interaction design should not be about simulating human communication but to engineer alternatives to the situated properties of interaction. Given the view of plans as event driven resources for action rather than as controlling structures the vagueness of plans is not a fault, but a consequence of the fact that intent and action must evolve side by side considering circumstantial and interactional particulars of specific situations. The foundation of actions is not plans but local interactions with our environment. The trick is to bring plans and particular circumstances into productive interaction.
Suchman concludes by stating that the project of building interactive machines has more to gain by understanding the differences between human interaction and machine operation, than by simply assuming their similarity and that the knowledge of these existing limitations should lead to new understanding regarding the design of machines as well as for understanding situated human action.
- Burton, R. & Brown, J. S. (1982). An investigation of computer coaching for informal learning activities. In Intelligent Tutoring Systems, D. Sleeman and J. S. Brown, eds. London: Academic Press.
- Cohen, P. & Perrault, C. R. (1979). Elements of a plan-based theory of speech acts. Cognitive Science 3:177-212.
- Suchman, L. A. (1987). Plans and Situated Actions: The problem of human-machine communication. Cambridge: Cambridge University Press.
- Weizenbaum, J. (1983). ELIZA: a computer program for the study of natural language communication between man and machine. Communications of the ACM, 25th Anniversary issue, 26(1):23-7. (Reprinted from Communications of the ACM, 29(1):36-45, January 1966.)