On Explaining Behavior

In his book Explaining Behavior, Fred Dretske distinguishes between types of representational systems based on whether the actions or intentions of an external agent need be mentioned in an explanation of the systemÕs behavior. He argues that the relevant criterion in making this distinction is whether the systemÕs behavior was structured by a learning process or by the activity of the systemÕs designer/constructor. However, our work with learning mobile robots demonstrates that the distinction ÒlearnedÓ versus ÒdesignedÓ: (i) is not always well-defined, and (ii) cannot by itself be used for deciding whether the activities of an outside agent must be mentioned in an explanation of a systemÕs behavior. We argue that DretskeÕs theory experiences similar difficulties even when the theory is applied to natural systems (e.g., humans). The problem arises because not all representational states fit neatly into one of the three idealized types he describes. There is a fairly broad class of states that fall somewhere between Type II and Type III. For system behavior involving these states, we ask the question anew: What is the relevant criterion for deciding whether the activities of an outside agent must be mentioned in an explanation of a systemÕs behavior? On Explaining Behavior Introduction In a series of books and articles, i Fred Dretske attempts to satisfy two pressing needs in the philosophy of mind by: providing a theory of content and demonstrating how content can be relevant in the explanation of behavior. Unlike most other philosophers writing on the same set of topics, ii Dretske has developed a theory and classification scheme that applies to all representational systems, not just those described as ÒpsychologicalÓ. While Explaining Behavior makes a major contribution to philosophersÕ understanding of content (especially psychological content), we are most interested in what Dretske has to say about explanation. Dretske distinguishes between triggering and structuring causes. A triggering cause is the event that caused some effect. (In the token causal process C _ E, C is the triggering cause.) A structuring cause is the cause of CÕs causing E (i.e., whatever was responsible for arranging things such that C caused E). Consider the case of a pigeon trained via instrumental conditioning to peck at a lever whenever a certain light goes on. At time t, the light goes on and the (hungry) pigeon pecks at the lever. The triggering cause of the pecking is the pigeonÕs seeing that the light is on. (This will be some state in the pigeonÕs brain.) The structuring cause of the pecking is the set of past training trials that ÒwiredÓ the pigeonÕs brain such that now (at time t) it pecks whenever it sees the light go on (and it is hungry). Note that structuring causal explanations are historical in nature: they advert to events that took place before the causing of the event to be explained. DretskeÕs contribution to the debate on the causal efficacy of mental content was in working out and then combining two separate parts of the puzzle: a theory of content and a description of the nature of content-adverting explanation. The first part of the puzzle was the development of an historical account of psychological content. According to this theory, psychological content is determined by the learning history of the system. Non-psychological content (the content had by non-learning representational systems) is determined by the intentions (and actions) of the designer/constructor of the representational system. Some philosophers of mind have argued that the physical properties of mental states Òscreen offÓ the mental properties (in particular, the content properties) of those mental states from playing a causal role. iii DretskeÕs second main contribution was noticing that this difficulty could be avoided by finding a role for content to play as structuring cause. The Type II/Type III distinction is a distinction between kinds of representational states. iv However, the distinction is also relevant to DretskeÕs theory of explanation. Consider again the causal process C _E. Assume that C is a representational state. If C is a Type II state, then the structuring cause of E will mention the actions of some outside agent. If, however, C is a Type III state, then the structuring cause of E will mention the learning history of the system that resulted in the C_ E connection. In the remainder of the paper, we focus on the Type II/Type III distinction as it applies to explanation. We begin by describing our work with learning mobile robots. We then argue that, at least according to one interpretation of the Type II/Type III distinction, our robot fits the bill for a Type III representational system. However, things are not quite so straightforward. We have noticed that there are serious difficulties in maintaining the Type II/Type III distinction in practice. We describe some problematic cases in the section entitled ÒBreakdown in the Type II/Type III DistinctionÓ. For example, it is possible to manipulate the learning process to the extent that the resulting representational states, while receiving their control duties as the result of a learning process, nevertheless have their content assigned by an outside agent. Suppose a heavily manipulated learning process results in a C _ E connection. Event C occurs, and causes E. It is unclear whether the structuring cause of E is the learning by the system or the manipulation by the outside agent (or, perhaps, some combination or the two). One should not be misled into thinking that this problem only arises for artificial learning systems; while it is more obvious for artificial systems (because artificial systems supply more avenues for manipulation), difficulties in classification as Type II or Type III (and concomitant difficulties in picking out structuring causes) arise for natural systems as well. The Robot and Some of What It Learns In this section, we provide information on our robot, its controller and (most importantly) the learning algorithm that runs Òon top ofÓ and influences the controller. We also try to give some idea of what the robot is capable of learning. This description of the robot is necessary, first, for establishing that the robot can be characterized as a Type III representational system, and, second, for giving the background necessary for understanding the difficulties that arise in applying the Type II/Type III distinction in practice. Our robot is a Pioneer-1 mobile robot, with several effectors and roughly forty sensors. vi The robot has learned numerous contingencies, including dependencies between its actions, the world state, and changes in the world state by processing data gathered as it roams around our laboratory. vii In this section, we focus on one learning method, clustering by dynamics, and a primitive ontology of actions that it learned without supervision. The robotÕs state is polled every 100msec., so a vector of 40 sensed values is collected ten times each second. These vectors are ordered by time to yield a multivariate time series. Figure 1 viii shows four seconds of data from just four of the PioneerÕs forty sensed values. Given a little practice, one can see that this short time series corresponds to the robot moving past an object. Prior to moving, the robot establishes a coordinate frame with an xaxis perpendicular to its heading and a y-axis parallel to its heading. As it begins to move, the robot measures its location in this coordinate frame. The ROBOT-X line is almost constant, while the ROBOT-Y line increases, indicating that the robot moved straight ahead. The VIS-A-X and VIS-A-Y lines indicate the horizontal and vertical locations, respectively, of the centroid of a patch of light on the robotÕs ÒretinaÓ, a CCD camera. VIS-A-X decreases, indicating that the objectÕs image drifts to the left on the retina, while VIS-A-Y increases, indicating the image moves toward the top of the retina. Then, both series jump to constant values simultaneously. ix In sum, the fourvariable time series in Figure 1 indicates the robot moving in a straight line past an object on its left, which is visible for roughly 1.8 seconds and then disappears from the visual field. Every time series that corresponds to moving past an object has qualitatively the same structure as the one in Figure 1 — namely, ROBOT-Y increases; VIS-A-Y increases to a maximum then takes a constant value; and VISA-X either increases or decreases to a maximum or minimum depending on whether the object is on the robotÕs left or right, then takes a constant value. ROBOT-X might change or not, depending on whether the robot changes its heading or not. It follows that if we had a statistical technique to group the robotÕs experiences by the characteristic patterns in time series, then this technique would in effect learn a taxonomy of the robotÕs experiences. Clustering by dynamics is such a technique. x The end result of applying clustering by dynamics to the multivariate sensor data gathered by the robot as it wanders around the lab is the learning of prototypes. A prototype is a characteristic pattern in the robotÕs time series of sensor data as it performs some action. The details of the clustering by dynamics algorithm are described in an endnote. xi In a recent experiment, this procedure produced prototypes corresponding to passing an object on the left, passing an object on the right, driving toward an object, bumping into an object, and backing away from an object. xii We claim that these prototypes were learned largely without supervision and constitute a primitive ontology of activities the robot learned some of the things it can do. What supervision or help did we provide? We wrote the programs that controlled the robot and made it do things. We divided the sensor data time series into episodes (although this can be done automatically). We limited the number of variables tha