Towards a unified theory of cognitive mapping

An integrated representation of large-scale space, or cognitive map, called PLAN is presented that attempts to address a broader spectrum of issues than has previously been attempted in a single model. Rather than examining wayfinding as a process separate from the rest of cognition, one of the fundamental goals of this work is to examine how the wayfinding process is integrated into general cognition. One result of this approach is that the model is "heads-up", or scene-based, because it takes advantages of the properties of the human visual system and particularly the visual system's split into two pathways. The emphasis on the human location or "where" system is new to cognitive mapping and is part of an attempt to synthesize prototype theory, associative networks and location together in a connectionist system. Not all of PLAN is new, however. Many of its parts have analogues in one or another pre-existing theory. What makes PLAN unique is the integration the various components into a coherent whole, and the capacity of this resulting system to speak to a wide range of constraints. Our approach emphasizes adaptiveness; thus our focus on such issues such as the ease of use and the efficiency of learning. The result is a model that has a stronger relationship both to the environment, and to the ways that humans interact with it, than previous models. The resulting model is examined in some detail and compared to other systems. PROTOTYPES, LOCATION, AND ASSOCIATIVE NETWORKS 3 Wayfinding is an important and complex task. The specialized structures that humans use for this task are called cognitive maps (Golledge, 1987). While cognitive maps are useful for a wide variety of reasons (Kaplan and Kaplan, 1982) their fundamental purpose is wayfinding. A cognitive map serves two functions with regard to wayfinding: representing environments, and the corresponding ability to use the representations to move from place to place within the mapped environments. This paper presents a representational theory of human cognitive mapping that addresses a wider spectrum of wayfinding issues than most other theories in the literature. Part of what makes this work unique in the cognitive mapping literature is that it draws on solutions from other domains of cognitive science and synthesizes them in a new way. In particular what stands out in PLAN (Prototypes, Location and Associative Networks) is an account of the impact that the capabilities of the human visual system, and in particular its two subsystems (the "what" and "where" systems), have on the kinds of cognitive maps that people ultimately develop. Human wayfinding can be broken down into four component problems: landmark identification, path selection, direction selection and creating abstract environmental overviews. Solutions to the first three of these problems are essential to human wayfinding. The fourth problem, achieving environmental abstraction, is a means of substantially increasing efficiency and functionality. While the four problems are individually separable, a unified solution requires an integrating framework to mesh them together. Landmark identification is the most basic component of wayfinding. Landmarks are environmental place markers vital in determining orientation and current location. The landmark identification problem in wayfinding is to separate out distinctive PROTOTYPES, LOCATION, AND ASSOCIATIVE NETWORKS 4 objects in the environment, called landmarks, which can later be used in route planning and can be recognized while traversing the chosen route. Path selection involves choosing a route to the goal. In this case a path is not a direction, but is more algorithmic, for example a series of places that will lead to the goal. In many models of cognitive mapping paths are conceptualized as sequences of landmarks. To follow a path one goes from landmark to landmark in the sequence; at each landmark in the path it will be necessary to select the next landmark in the sequence. Direction selection involves choosing a direction in which to travel. If the goal is in sight, for example, a reasonable direction to pick would be towards the goal. For goals which are not in sight the direction selection problem is of course more difficult; beyond the fact that the goal cannot be seen, sometimes a journey will require a series of turns and shifts in direction. Thus direction selection at the starting point is rarely sufficient to guide an entire route. Abstract environmental overviews are a further generalization of the route concept. If one were to travel extensively in a particular environment it would be useful to have a coherent overview of the entire environment. Rather than dealing with routes individually, such a structure would allow them to be extracted from a common abstraction. In addition this overview would make large-scale reasoning about the environment simpler. However, while these overviews do serve to increase the efficiency of wayfinding, they are not strictly necessary. Thus we will include a discussion of wayfinding without such capabilities.environmental overviews are a further generalization of the route concept. If one were to travel extensively in a particular environment it would be useful to have a coherent overview of the entire environment. Rather than dealing with routes individually, such a structure would allow them to be extracted from a common abstraction. In addition this overview would make large-scale reasoning about the environment simpler. However, while these overviews do serve to increase the efficiency of wayfinding, they are not strictly necessary. Thus we will include a discussion of wayfinding without such capabilities. Each of these four subproblems has a different character and consequently is likely to require a different solution. The landmark identification problem primarily concerns the PROTOTYPES, LOCATION, AND ASSOCIATIVE NETWORKS 5 object recognition system. The path selection problem is cognitive, often requiring the selection of one path out of a number of alternatives. The direction selection problem, while generally visual, is more locational than the landmark identification problem. Finally, the problem of creating an abstract environmental overview requires a hierarchical synthesis of each of the other three solutions. While we are taking a modular approach, it need not be the case that the subsystems are completely separated; for example, knowing what to expect next on a path may be a useful bit of information when indentifying a landmark. Therefore, while we will discuss each of the problems separately, throughout the paper we will be careful to note how the different parts interact in the overall theory. The solutions offered by PLAN to each of these problems are not necessarily new. Indeed, this model builds upon our own previous work, and that of the connectionistoriented SESAME1 group. In particular, PLAN represents a synthesis of prior work on associative networks (Kaplan, 1973; Levenick, 1991), on prototypes (Kaplan, et al., 1991), and on spatial or locational visual processing (Lesperance, 1990). Thus what is new is the attempt to construct a coherent synthesis of several previously divergent lines of research. The complementarity of these various components has, in fact, exceeded our expectations. The resulting system has a number of interesting properties, among them economy, simplicity and error tolerance. This paper begins by laying out the key constraints that have driven the PLAN model. We take for granted that PLAN is a connectionist model; since the portions of our previous work incorporated in PLAN are all connectionist, this constraint is implicit. (Readers interested in the SESAME approach to connectionist modelling can refer to (Kaplan, et al., 1990; Kaplan, et al., 1991) The remaining constraints can be divided into three categories: 1) Constraints deriving from the fact that cognitive maps were developed in an evolutionary context. 2) Physiological constraints, particularly pertaining to the brain PROTOTYPES, LOCATION, AND ASSOCIATIVE NETWORKS 6 structure associated with processing spatial information. 3) What we will call developmental constraints, but which will also subsume general psychological constraints derived from studies of human behavior. In the third section of the paper we will present the PLAN model. While we will compare specific parts of PLAN to pieces of other models along the way, we save extended comparisons until after the entire model has been presented. The fourth section, therefore will center around a more general discussion of PLAN and how it compares to and differs from other theories in the literature.