On Predicting Behavior
Our ability to predict and explain the behavior of other humans has long been of interest for both philosophers and psychologists. Initially, it was taken for granted that this competence is based on an ability to attribute beliefs and desires to other humans. For example, Daniel Dennett's (1987) intentional stance, which requires that intentional behavior is best predicted via belief attributions, rests on this assumption. Premack and Woodruff (1978) also rely on this view in stating that if a chimpanzee can predict intentional behavior, then he has a theory of mind (ToM); that is, he can attribute beliefs and desires to others. These assumptions as to the nature of our predictive ability have developed into a general theory of the mind. It is thought that if we make use of beliefs and desires to predict behavior, then we must have a tacit-theory of the psychological influences on behavior (Stich and Nichols 1996, Fodor 1987, Wellman 1990). In contrast to this tacit-theory view, it has been suggested (e.g. Harris 1990, Goldman 1995, Gordon 1995) that we need not appeal to others' beliefs and desires in order to predict behavior, and instead we simulate by, roughly, putting ourselves in the position of another. I argue that the apparent dichotomy between these two theories is not actual. There is a third method we commonly use to predict behavior, and this, conjoined with the above two methods, goes a long way to describing how we are able to predict the huge number of human and non-human intentional behaviors.
The tacit-theory theory has arisen from much of the recent psychological interest in ToM. If one is able to entertain a propositional attitude such as "Sue believes that the kettle is hot" then she has a ToM. This belief of Sue's can be used to predict her behavior: that she won't touch the kettle without a mitt, or that she will now make tea, etc. ToM is not to be confused with a philosophical theory as to the nature of mind and mental states; "theory of mind" as understood broadly only refers to our folk understanding of others minds as entertaining beliefs and desires different from our own. This ability is sometimes called folk, or belief-desire psychology.
Though interest in ToM was introduced via a research strategy within comparative psychology and cognitive ethology (Premack and Woodruff 1978), interest in it has spread widely from its initial appearance within the context of animal research. Over the past several years, ToM has become more widely studied in developmental psychology, and the results have far exceeded those of cognitive ethology or comparative psychology. The study of ToM in humans greatly expanded after the introduction by Wimmer and Perner (1983) of a test for ToM in children. They developed what is known as a false-belief task, which tests a child's ability to predict behavior via belief attributions.
Leslie (1991) gives an example of such a test in which the subjects are shown a puppet show. There are two puppet actors, Sally and Anne. Sally hides a marble in a basket, then leaves the room. While she is gone, Anne takes the marble from the basket and hides it in a box. Then Sally returns to the room, and the subject is asked where Sally will look for her marble. A subject who reports that Sally will look for the marble in the basket, where Sally last thought it was, is said to have a ToM. However, if the subject says that Sally will look for the marble in the box, then he has not shown the ability to distinguish his own beliefs from the beliefs of others, and thus he is thought not to have a ToM. Normal children of about four years typically pass the false-belief task, predicting that Sally will look in the basket, while younger children fail it.
The philosopher's involvement in the ToM research has been mostly at the level of developing metarepresentational theories consistent with ToM. The typical explanation is that within the human cognitive architecture there is something analogous to a database of conditional rules describing the interaction of behavior, knowledge and motivation of agents (Stich and Nichols 1996). These rules can then be used to predict and explain intentional behavior. For example, one such rule may be: "If someone drops his wallet and knows he dropped his wallet, then, all things being equal, he will pick it up." The database would also include information about motivation: "He will pick up his wallet because, all things being equal, people want to keep things that are useful to them." With this sort of knowledge those with a fully developed ToM can predict the future behavior of agents to whom these generalizations are applicable, and in addition we can explain their behavior by referring to the relevant motivational rule.
The simulation theory has arisen in direct response to the rule-based account of our cognitive architecture as advocated by the tacit-theorists. Prediction of behavior via simulation does not require belief or desire attribution. Instead of appealing to tacit rules, a person simulates what it is like to be another person and from that she can predict what the other would do by first determining what she herself would do in that situation. According to simulation theories, we are able to predict and explain the behavior of others through use of our own cognitive processes. The success of this strategy is based on an assumption that in similar situations and with similar beliefs, people will act in similar ways. So, for the simulation theorists, knowledge of others' mental states requires inputting the stimulus of the subject's situation to one's own cognitive architecture conjoined with an inference that the subject's belief-structure is similar to one's own (Harris 1991, 1992).
Mental simulation is seen as analogous to non-cognitive simulations. In general, future events can be predicted with some degree of certainty if all past similar situations have led to the same result. For example, take the case of car air-bags. The air-bags in my car have never been inflated, though I expect that they will inflate if the car should impact something with a certain force, because the car manufacturers have tested other cars very similar to mine in simulated crashes. These test-crashes have given me the evidence I need to be able to make predictions regarding my car, even if I don't know anything about the inner workings of air-bags. I need have no theory or knowledge regarding the physics on which air-bag inflation rests; instead I use an exemplar, or a set of exemplars, and extrapolate from them by assuming that my car is similar to the others. In the case of mental simulation, the exemplar is oneself, and an analogy is drawn to other people. If I want to predict what a person would do if he dropped his wallet, I would first ask myself what I would do if I dropped my wallet. We imagine ourselves in the person's position ("Wouldn't it be terrible if I lost my wallet?!") and then generalize from analogy that the person would want to pick up the wallet as well.
Though the debate between simulation and tacit-theory is still raging in psychology and philosophy, there has been growing support for a hybrid position which accepts that some of our predictions are made through belief and desire attribution, and others are made through simulation (Kuehberger and Perner 1997). In this paper I wish to embellish on this view of how we predict behavior. I will do this through examining two sets of questions which are not sufficiently addressed within this debate:
An analysis of these questions will suggest that there is a third method we use to predict behavior, thus showing that tacit-theory and simulation are not the only methods one might use. This method, like simulation, does not require the attribution of belief or desire, though it may be seen to require something akin to general rules, as tacit-theory does. I will begin to defend the existence of such a method by examining the first of my two sets of questions.
In order to determine how we gain knowledge about specific human behavior, let's look at the example of someone dropping a wallet. As young children, we would not predict that the wallet would be picked up if it were noticed on the ground, for young children have not yet learned the value of a wallet, nor have they much experience with people dropping wallets. How do children learn that a dropped wallet will be picked up? Certainly, children learn many things without explicit instruction, language being a paradigm example. Most children probably learn through observation that dropped wallets are usually picked up. They may observe some incidents where wallets are dropped or misplaced, and recognize that the owner of the wallet either picks it up, or otherwise demonstrates that they wish to be in possession of the wallet. We could say that a child develops an implicit theory that people value wallets in general, as the tacit-theory theory suggests. Additionally, a child who does not yet recognize any desire for a wallet, having no personal use for credit cards, a driver's license, etc., may still be able to predict that a dropped wallet will be picked up after having recognized the value adults place on wallets. So, were a child to make a prediction that an adult would pick up a wallet, when the child herself has no desire for one, her ability to predict might initially seem to lend support to the tacit-theory theory rather than the simulation theory. Pure simulators claim that there is no belief/desire attribution involved at any level in predicting behavior. They avoid an initial attribution of belief or desire in setting up the simulation by claiming that knowledge of the external situation is enough to simulate well (Gordon 1995). If the child has no need for the wallet, and does not appeal to the adult's different beliefs and desires regarding wallets, then no amount of knowledge of the immediate environment will enable her to make the correct prediction. If one is able to predict behavior of someone with different desires, pure simulation cannot be the explanation.
However, this conclusion does not lend support for tacit-theory, for there does not seem to be any need for the child to know that the adult has a belief or desire in order to predict that the wallet will be picked up; she need not attribute any intentions to the adult in order to make her prediction. Instead, the child may have been led to expect that a dropped wallet will be picked up, because dropped things are generally picked up by adults. This sort of reasoning is what makes up our folk physics, and many of our everyday predictions are inductive generalizations from past experience.
The generalization "If someone drops a wallet, then she will pick it up" can be thought of as an implicit rule that one uses to predict behavior, but it involves no attribution of belief or desire. ToM, on the other hand, takes humans to predict the behavior of others through belief and desire attribution. So, there seem to be some varieties of prediction which require neither belief attribution nor simulation, and these are predictions gained through induction.
In order to form any hypothetical database, and in order to know what one would do oneself in a given situation, many of these sorts of rules must be gained through induction. Though the development of children involves learning many very basic rules, this process never stops, for our experience continues to provide us with new instances on which to generalize as the variety of our experiences expand.
As an answer to the question, How are the rules of the database formed?, the preceding remarks suggest that the rules are primarily formed by inductive generalizations from observations, which need not require any knowledge or supposition of the beliefs or desires of an agent. The remaining question for the tacit-theory theorists is how one adjudicates between two or more rules when in a more complex situation. For example, suppose a child observes someone dropping a wallet as he notices a charging elephant. There are at least two potential salient characteristics of the situation-the dropped wallet and the elephant. To predict the behavior of this person using the tacit-theory, the child would then activate the dropped wallet rule, as well as the rule regarding being chased by wild animals. That rule may be something like this: "If one is being chased by a wild animal, then, all things being equal, he will try to get away from the animal." The problem for the tacit-theory theorist here is how we recognize that the wild animal rule will trump the dropping the wallet rule (for I predict that most people would ignore the dropped wallet in such a situation). There must be, in addition to these basic sorts of rules, further second-order rules regarding relative values, salient characteristics, and adjudication between conflicting rules. This sort of problem is notorious within ethics and dilemma theory, and particularists within ethics criticize rule-based ethical theories for being unable to resolve such questions. I won't attempt to answer such problems here; I only wish to point out that as the system of rules within our tacit theory is examined closely these problems arise and they must be dealt with if we maintain that the tacit-theory is to be seen as the way we predict behavior. (1)
The simulation theorist faces another set of questions, including: How is it that we are able to predict our own behavior? On this topic simulation theorists have little to say, though it seems that an answer to this is essential for simulation to be a plausible theory. In order to use oneself as the exemplar, one needs some method of predicting her own behavior before she can to generalize to others. This is a question regarding the structure of our cognitive mechanisms. As such, it presents a scenario in which tacit-theory and simulation are compatible.
For example, one possible way of predicting our own behavior is through an appeal to a tacit theory. In this case, the database might include rules specific to oneself, such as "If I drop something, then all else being equal I will pick it up" and so forth. We might appeal to our own database to anticipate how we will behave in certain circumstances, though this database is personal and would not generalize to other people. Instead, to predict what another person will do, we first pretend to be in that situation ourselves and appeal to the database in order to determine what we would do, and only then could we generalize to the third party. It is highly unlikely that any pure simulation theorist would accept this account of self-prediction, for they want to avoid the notion of a database all together. Indeed, one of the strengths of simulation theory is that it is seen as not needing such a database (Stich and Nichols 1996).
However, if the above theory of how we predict our own behavior is false, then another account much be provided. It will not do to leave a black box within our cognitive architecture where predictions of one's own behavior occur. Nor will it do to say that we just know how we will behave. It certainly seems to us most of the time that we do just know, but this lends no more support to simulation than it does to tacit-theory. Tacit-theory does not require that we have an overt experience of searching though a database looking for rules; that is why the theory is tacit. Tacit-theory offers an explanation of how we predict our own and other's behavior, while simulation theory leaves out the method of self-prediction, on which the other-prediction rests.
Through examining how we learn how to behave as children, it becomes clear that more is involved in our prediction of behavior than merely simulation or tacit-theory. When children are being socialized into a community, they are given implicit and explicit reinforcement to behave in certain ways which are considered normal by the community. We learn to do x in situation y and not in z, and we learn to follow many such rules not by formulating the rules for ourselves, but through following what becomes a natural inclination. As socialized creatures, we respond to our emotions of guilt and acceptance through our behavior. After having been trained, we feel bad if we do x in z rather than y, and since humans don't like to feel bad, we avoid such behavior. Such responses can be described and explained in a conditional rule, as I have done above, but from that it does not follow that we need to appeal to such rules in order to predict either our own behavior or the behavior of others.
Let's look at another example. When I go to buy a cup of coffee at an American coffee house, I walk up to the counter, ask for a cup of coffee, and expect that I will be asked for some amount of money and given a cup of coffee. Having done this thousands of times, I have formed expectations without needing to attribute beliefs to the person behind the counter, and I need not imagine what it would be like for me to be that person. In such familiar situations, I know that the person will behave in a certain way using the same method of prediction I use when I expect that my car will start when I turn the key in the ignition. And I know it for the same reasons: we learn to expect a particular action in a particular circumstance after having experienced many cases of the conjunction. I have no grand theory of the inner workings of my car, and I need no grand theory of the human mind to predict that I will be asked for a dollar and given a cup of coffee. In most of our daily interactions with people on the street, we are in a continual state of anticipating behavior. When walking down a city street we are confronted with hundreds of people with whom we interact, if even for a second when moving aside to allow someone to pass. Our verbal and non-verbal interaction with other people is a constant dance choreographed through our society and our human impulse to anticipate similarity. To explain the individual moves of this intricate dance we appeal to beliefs and desires, but for many of the simple everyday predictions we make, neither belief attribution nor simulation is necessary.
Of course, the induction account is not exhaustive. We certainly do sometimes appeal to our own behavior in order to predict the behavior of others. We do say to ourselves, not infrequently: "What would I do if I were him?" And we also frequently predict behavior by attributing beliefs and desires, especially in complex or unusual circumstances. E.g., a realtor who is trying to sell a house may try to list the desires and beliefs of the customer in order to make the most convincing case for buying that particular house. There are many different kinds of behavior which we are capable of predicting to different degrees. It is easier to predict the behavior of the barista in the coffee house than it is to predict what she would do when confronted with a moral dilemma. It is often easier to predict what one's family and friends will do than to predict what strangers would do in the same situation. And it is not always easier to predict the behavior of someone similar to oneself than someone dissimilar. For, if I saw a racist in a position where she could save only one of two drowning children equidistant from her, I would predict that she would save the child of her own race, and leave the other. If I placed myself in the same situation, I have little faith in any self-prediction I could now make.
There is no denying that we do use both belief/desire attribution and simulation to predict the behavior of other humans. The point of this paper is to acknowledge that there is not a dichotomy between these two positions, since we predict via other methods as well. Sometimes it is simpler (as in the case of the coffee house) to use the inductive method than either of the other two, and in some cases it will be necessary to use the inductive method. For example, if I am trying to predict the behavior of someone utterly unlike me, I will not know what his beliefs and desires are, and I will need to study the subject for a while in order to determine what his behaviors are. Simulation theory should only work with agents who are relevantly similar to oneself, though we can become quite good at predicting the behavior of other agents. We often predict the behavior of non-human animals who are quite different from ourselves, and we can do this after observing them and determining how they actually do behave.
There are many ways of predicting behavior, and it seems that, of the three methods of prediction discussed in this paper, all are valuable tools which we do utilize. This paper has presented two ideas which require further investigation. I have suggested that there may be no conflict between simulation and tacit-theory, since it may be discovered that the self-prediction necessary within simulation theory is made through appeal to tacit theories. The second point is foundational to these issues, and requires independent investigation. That is the suggestion that predicting behavior is quite different from explaining behavior.
The main conclusion to be drawn from this paper is there is at least one other method that we use to predict behavior, the inductive strategy. Prediction that appears to be inconsistent with the tacit-theory approach should not assumed to be produced through simulation by default, and vice versa, because there is another possible explanation.
(1) I see this as a problem with the characterization of tacit-theory as an actual theory. The same problem can be raised for the inductive strategy I am advocating adding to the mix. However, the inductive strategy is rule-based only if the rest of our cognitive architecture is rule-based, and the last decade of artificial intelligence research suggests that this is not the case. I propose that a connectionist cognitive architecture would avoid the characterization of belief/desire psychology as rule-based, and instead, like our inductive knowledge, psychological knowledge would be represented in a decentralized, untheoretical fashion. In this manner one may be able to avoid the above criticism.
Dennett, Daniel. 1987. The Intentional Stance, MIT Press.
Fodor, J. A. 1987. Psychosemantics: The Problem of Meaning in the Philosophy of Mind. MIT Press.
Goldman, Alvin. 1995. Interpretation psychologized. In Martin Davies and Tony Stone, eds. Folk Psychology. Blackwell Publishers.
Gordon, Robert. 1995. Folk psychology as simulation. In Martin Davies and Tony Stone, eds. Folk Psychology. Blackwell Publishers.
Harris, Paul. 1992. From simulation to folk psychology: The case for development. Mind and Language 7: 120-144.
Harris, Paul. 1991. The work of the imagination. In Andrew Whiten, ed. Natural Theories of Mind: Evolution, Development, and Simulations of Everyday Mindreading, Basil Blackwell.
Kuehberger, Anton, and Perner, Josef. 1997. Predicting others through simulation or by theory? A method to decide. Presented at the European Society for Philosophy and Psychology Annual Meeting, Padova Italy.
Leslie, Alan. 1991. The theory of mind impairment in autism: Evidence for a modular mechanism of development? In Andrew Whiten, ed. Natural Theories of Mind: Evolution, Development, and Simulations of Everyday Mindreading, Basil Blackwell.
Premack, David and Woodruff, G. 1978. Does the chimpanzee have a theory of mind? In Behavioral and Brain Sciences 1: 515-26.
Stich, Stephen and Nichols, Shaun. 1996. How do minds understand minds? Mental simulation versus tacit theory. In Stephen Stich Deconstructing the Mind. Oxford University Press.
Wellman, Henry. 1990. The Child's Theory of Mind. MIT Press.
Wimmer, H. and Perner, J. 1983. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition 13: 103-128.