Use this URL to cite or link to this record in EThOS:
Title: A computational inquiry into navigation, with particular reference to the hippocampus
Author: Foster, D. J.
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 1999
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis reviews evidence that a neural structure in the mammalian brain known as the hippocampus provides a representation of space that is purely local. The thesis then reviews a selection of computationally efficient, neurally plausible, reinforcement learning methods that can use local representation to learn optimal actions in navigation-like tasks. The key disadvantages usually associated with the methods are slow learning, and inflexibility to change. Both disadvantages were investigated in this thesis in the context of learning to navigate. A first model is presented which learns to perform two hippocampally dependant tasks, one involving navigation to a single goal, the other involving navigation to multiple goals. Animals gradually acquire the single goal tasks, but in the multiple gaols task, gradually acquire the ability to navigate directly to a novel goal on only the second trial to that goal. One component of the model uses place cells within a standard reinforcement learning scheme, temporal difference learning in an actor-critic. This component by itself captures performance in the single goal task, but fails to capture one-trial learning in the multiple goals task. A second component of the model learns globally consistent co-ordinates from local self-motion information, in a novel application of temporal difference learning. This co-ordinate learning is relatively independent of the behaviour of the animal, enabling the gradual acquisition of one-trial learning in the multiple goals task to be captured. Two purely behavioural predictions follow which were tested experimentally. First, once a co-ordinate system has been learned, simple placement of an animal at a novel goal should provide the animal with sufficient information to allow direct paths on the next trial, and this was shown to be true. Second, since co-ordinates often no principled way of circumnavigating barriers, the model predicts that animals will be unable to learn the task in the presence of barriers, but this was shown to be false.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available