USC Brain Project: Hippocampus and Navigation Group

The TAM-WG Model: WG Influence

-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
 

Theoretical background:

In the WG theory, there is an edge from node x to node x' in the graph for each distinct and direct path the animal has traversed between a place it recognizes as x and a place it recognizes as x'. This is implemented in the WG model as a link between two nodes of the World Graph layer.

Edge information allows the animal to perform goal-oriented behaviors. In the WG theory, this is obtained by incorporating expected drive reduction information to each node of the world graph in the form of a vector. The WG model, however, is able to learn expectations of future reward. As in TAM, this is implemented by the use of temporal differences learning in an actor-critic architecture (Barto et al., 1983; Barto, 1994; Sutton, 1988). In the WG model, however, expectations of future reinforcement are associated with pairs of nodes/edges, and not with state/actions as is the case in TAM. For this reason, when a particular node of the world graph is active and, for example, the simulated animal is hungry, all it needs to do is to select the edge containing the biggest hunger reduction expectation and follow its direction towards the node it points to.

For more information on reinforcement learning or on the WG theory, please refer to the cited papers or to the papers listed in the Hippocampus and Navigation Group homepage.

-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

BSL Home PageUSC Home Page

University of Southern California Brain Simulation Lab

All contents copyright (C) 1994-1997.
University of Southern California Brain Simulation Lab
All rights reserved.
Author: Alex Guazzelli <aguazzel@rana.usc.edu>