Short and medium range navigation and its relationship to cognitive mapping and associative learning

• Main Author: Biegler, Robert
• Published: University of Edinburgh 1996
• Subjects: 153
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.712216

Learning allows organisms to predict and prepare for events in the environment that are not sufficiently regular that responses to each situation could be genetically hardwired. A possible categorization of what can be learned is as follows: First, an animal may learn that an event is likely to happen. This means learning predictive relations between events, or the probability that an event A occurs with an event B, rather than independently. Second, they may form a representation of the magnitude of the event. Third, animals may learn when an event is likely to happen, the temporal relations between events. Fourth, they may acquire knowledge where something will happen, the spatial relations between events. The question arises whether these distinctions are merely convenient labels or reflect genuine differences between dissociable psychological variables and perhaps processes. The most widely accepted account of animal learning, associative learning theory, assumes that information from all these variables is collapsed into only a single output variable: the strength of an associative link. The theoretical framework of associative learning has predominantly been developed and tested within the domain of learning about predictive relationships between events, weighted by event relevance. The requirements for navigation through space are in some respects quite different. Animals can influence the rate and direction of their passage through space. In the two or three dimensions of space shortcuts and detours become possible. The computation of path length may require vector addition. Possible goals of computation will be considered and compared to data on the contents, acquisition and manipulation of spatial representations. The experimental part of this thesis concentrates on two aspects of information acquisition, landmark stability and blocking. Animals appear to weigh information from different sources according to two different and normally opposed criteria, accuracy and reliability. If discrepancy between two such sources is small, more weight will be given to the more accurate source of information, if the discrepancy is large more to the reliable source. The experiments on landmark stability suggest that manipulating discrepancy throughout training will influence animals' estimate of reliability of a source of information. Other manipulations of this estimate, independent from discrepancy, are also possible. The experiments on blocking have not yielded a simple result. Blocking occurs when the animals were trained with one of two landmark arrays; the other array led to an enhancement of performance when testing with the added landmarks. In addition, previous work on the "geometric module" has been extended and a novel weighting of landmarks by position in the array, rather than distance from a goal, has been found. It is argued that there associative learning can play a role in the creation of most possible representations of space, but that some aspects of navigation involve computations which associative learning is not capable of. Further, consideration of the possible functions of navigation suggests that there is no clear dichotomy between mapping and non-mapping strategies. The features of cognitive maps derived from analogy with physical maps do not form an indivisible category. A navigational system may have only some of these properties, depending on what is required of it. The supposed incompatibility of cognitive mapping and associative learning does not exist, both because there are several navigational strategies that could be considered cognitive mapping processes and because associative learning could contribute to most of them.